diff --git "a/parse/dev/fcO9Cgn-X-R/fcO9Cgn-X-R_middle.json" "b/parse/dev/fcO9Cgn-X-R/fcO9Cgn-X-R_middle.json"
new file mode 100644--- /dev/null
+++ "b/parse/dev/fcO9Cgn-X-R/fcO9Cgn-X-R_middle.json"
@@ -0,0 +1,39431 @@
+{
+ "pdf_info": [
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 120,
+ 97,
+ 492,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 493,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 493,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Capturing Failures of Large Language Models via",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 213,
+ 116,
+ 399,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 213,
+ 116,
+ 399,
+ 139
+ ],
+ "score": 1.0,
+ "content": "Human Cognitive Biases",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 160,
+ 179,
+ 266,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 188,
+ 177,
+ 238,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 188,
+ 177,
+ 238,
+ 192
+ ],
+ "score": 1.0,
+ "content": "Erik Jones",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 185,
+ 189,
+ 241,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 189,
+ 241,
+ 203
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 159,
+ 202,
+ 267,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 202,
+ 267,
+ 214
+ ],
+ "score": 1.0,
+ "content": "erjones@berkeley.edu",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 324,
+ 179,
+ 451,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 349,
+ 178,
+ 425,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 349,
+ 178,
+ 425,
+ 190
+ ],
+ "score": 1.0,
+ "content": "Jacob Steinhardt",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 359,
+ 188,
+ 416,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 359,
+ 188,
+ 416,
+ 204
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 322,
+ 201,
+ 452,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 201,
+ 452,
+ 213
+ ],
+ "score": 1.0,
+ "content": "jsteinhardt@berkeley.edu",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 240,
+ 328,
+ 254
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 240,
+ 330,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 240,
+ 330,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 267,
+ 469,
+ 420
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 267,
+ 471,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 267,
+ 471,
+ 279
+ ],
+ "score": 1.0,
+ "content": "Large language models generate complex, open-ended outputs: instead of out-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 291
+ ],
+ "score": 1.0,
+ "content": "putting a class label they write summaries, generate dialogue, or produce working",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 469,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 469,
+ 302
+ ],
+ "score": 1.0,
+ "content": "code. In order to asses the reliability of these open-ended generation systems, we",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 140,
+ 298,
+ 470,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 298,
+ 470,
+ 313
+ ],
+ "score": 1.0,
+ "content": "aim to identify qualitative categories of erroneous behavior, beyond identifying",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 311,
+ 469,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 311,
+ 469,
+ 322
+ ],
+ "score": 1.0,
+ "content": "individual errors. To hypothesize and test for such qualitative errors, we draw",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 321,
+ 470,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 321,
+ 470,
+ 333
+ ],
+ "score": 1.0,
+ "content": "inspiration from human cognitive biases—systematic patterns of deviation from ra-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 333,
+ 469,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 333,
+ 469,
+ 344
+ ],
+ "score": 1.0,
+ "content": "tional judgement. Specifically, we use cognitive biases as motivation to (i) generate",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "score": 1.0,
+ "content": "hypotheses for problems that models may have, and (ii) develop experiments that",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "score": 1.0,
+ "content": "elicit these problems. Using code generation as a case study, we find that OpenAI’s",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 365,
+ 469,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 365,
+ 469,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Codex errs predictably based on how the input prompt is framed, adjusts outputs",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 388
+ ],
+ "score": 1.0,
+ "content": "towards anchors, and is biased towards outputs that mimic frequent training exam-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 386,
+ 470,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 386,
+ 470,
+ 399
+ ],
+ "score": 1.0,
+ "content": "ples. We then use our framework to elicit high-impact errors such as incorrectly",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 397,
+ 469,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 397,
+ 469,
+ 410
+ ],
+ "score": 1.0,
+ "content": "deleting files. Our results indicate that experimental methodology from cognitive",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 409,
+ 423,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 409,
+ 423,
+ 421
+ ],
+ "score": 1.0,
+ "content": "science can help characterize how machine learning systems behave.1",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 15.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 441,
+ 191,
+ 455
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 192,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 192,
+ 457
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 467,
+ 504,
+ 501
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 479
+ ],
+ "score": 1.0,
+ "content": "Recent large language models have achieved new, exciting capabilities. In contrast to traditional",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 490
+ ],
+ "score": 1.0,
+ "content": "classifiers, these models can generate open-ended text, enabling use cases like summarization",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 108,
+ 488,
+ 485,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 488,
+ 485,
+ 501
+ ],
+ "score": 1.0,
+ "content": "[Stiennon et al., 2020], dialog [Thoppilan et al., 2022], and code generation [Chen et al., 2021]",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 506,
+ 505,
+ 582
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "The open-ended power of these systems, however, poses new reliability challenges. We must",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 528
+ ],
+ "score": 1.0,
+ "content": "understand not only when systems err, but also the kinds of errors they make, as some errors are much",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "more costly than others. For example, erroneous code that does not compile is less dangerous than",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "score": 1.0,
+ "content": "code that deletes all files in the home directory. Studying how frequently an error occurs is difficult,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 506,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 506,
+ 561
+ ],
+ "score": 1.0,
+ "content": "as the same error (e.g. delete all files) can appear in a wide range of syntactically diverse outputs.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "In order to better reason about how complex systems err, we need methods to test whether systems",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 495,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 495,
+ 584
+ ],
+ "score": 1.0,
+ "content": "make the same qualitative error across different prompts, even when the generated outputs differ.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 587,
+ 505,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "To study these reliability challenges, we primarily focus on code generation models. Such models",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "complete programs from comments, descriptions of code functionality, or initial lines of code.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "Code generation is particularly amenable to study since it is objective: generated solutions are",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 505,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 505,
+ 632
+ ],
+ "score": 1.0,
+ "content": "unambiguously correct or incorrect. Yet it is also open-ended: the set of programs a model could",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 506,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 506,
+ 643
+ ],
+ "score": 1.0,
+ "content": "output is arbitrarily large, so the rate at which a specific program is outputted is not very descriptive.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 647,
+ 505,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "Many of the reliability challenges posed by code generation models, and open-ended systems broadly,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "also arise when studying qualitative failures in human decision making. These failures, called",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 669,
+ 506,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 669,
+ 506,
+ 682
+ ],
+ "score": 1.0,
+ "content": "cognitive biases, are systematic ways in which humans deviate from rational judgment [Tversky",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 693
+ ],
+ "score": 1.0,
+ "content": "and Kahneman, 1974]. For example, Tversky and Kahneman find that humans inadequately adjust",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5
+ }
+ ],
+ "page_idx": 0,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 106,
+ 731,
+ 385,
+ 742
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 730,
+ 385,
+ 743
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 730,
+ 385,
+ 743
+ ],
+ "score": 1.0,
+ "content": "36th Conference on Neural Information Processing Systems (NeurIPS 2022).",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 118,
+ 701,
+ 459,
+ 713
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 462,
+ 715
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 462,
+ 715
+ ],
+ "score": 1.0,
+ "content": "1 Code for this paper is available at https://github.com/ejones313/codex-cog-biases.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 120,
+ 97,
+ 492,
+ 137
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 493,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 120,
+ 97,
+ 493,
+ 119
+ ],
+ "score": 1.0,
+ "content": "Capturing Failures of Large Language Models via",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 213,
+ 116,
+ 399,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 213,
+ 116,
+ 399,
+ 139
+ ],
+ "score": 1.0,
+ "content": "Human Cognitive Biases",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 160,
+ 179,
+ 266,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 188,
+ 177,
+ 238,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 188,
+ 177,
+ 238,
+ 192
+ ],
+ "score": 1.0,
+ "content": "Erik Jones",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 185,
+ 189,
+ 241,
+ 203
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 185,
+ 189,
+ 241,
+ 203
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 159,
+ 202,
+ 267,
+ 214
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 159,
+ 202,
+ 267,
+ 214
+ ],
+ "score": 1.0,
+ "content": "erjones@berkeley.edu",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 159,
+ 177,
+ 267,
+ 214
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 324,
+ 179,
+ 451,
+ 213
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 349,
+ 178,
+ 425,
+ 190
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 349,
+ 178,
+ 425,
+ 190
+ ],
+ "score": 1.0,
+ "content": "Jacob Steinhardt",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 359,
+ 188,
+ 416,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 359,
+ 188,
+ 416,
+ 204
+ ],
+ "score": 1.0,
+ "content": "UC Berkeley",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 322,
+ 201,
+ 452,
+ 213
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 322,
+ 201,
+ 452,
+ 213
+ ],
+ "score": 1.0,
+ "content": "jsteinhardt@berkeley.edu",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 322,
+ 178,
+ 452,
+ 213
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 283,
+ 240,
+ 328,
+ 254
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 281,
+ 240,
+ 330,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 281,
+ 240,
+ 330,
+ 255
+ ],
+ "score": 1.0,
+ "content": "Abstract",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 142,
+ 267,
+ 469,
+ 420
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 141,
+ 267,
+ 471,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 267,
+ 471,
+ 279
+ ],
+ "score": 1.0,
+ "content": "Large language models generate complex, open-ended outputs: instead of out-",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 291
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 277,
+ 470,
+ 291
+ ],
+ "score": 1.0,
+ "content": "putting a class label they write summaries, generate dialogue, or produce working",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 141,
+ 288,
+ 469,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 288,
+ 469,
+ 302
+ ],
+ "score": 1.0,
+ "content": "code. In order to asses the reliability of these open-ended generation systems, we",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 140,
+ 298,
+ 470,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 140,
+ 298,
+ 470,
+ 313
+ ],
+ "score": 1.0,
+ "content": "aim to identify qualitative categories of erroneous behavior, beyond identifying",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 141,
+ 311,
+ 469,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 311,
+ 469,
+ 322
+ ],
+ "score": 1.0,
+ "content": "individual errors. To hypothesize and test for such qualitative errors, we draw",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 141,
+ 321,
+ 470,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 321,
+ 470,
+ 333
+ ],
+ "score": 1.0,
+ "content": "inspiration from human cognitive biases—systematic patterns of deviation from ra-",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 141,
+ 333,
+ 469,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 333,
+ 469,
+ 344
+ ],
+ "score": 1.0,
+ "content": "tional judgement. Specifically, we use cognitive biases as motivation to (i) generate",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 343,
+ 470,
+ 356
+ ],
+ "score": 1.0,
+ "content": "hypotheses for problems that models may have, and (ii) develop experiments that",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 354,
+ 470,
+ 367
+ ],
+ "score": 1.0,
+ "content": "elicit these problems. Using code generation as a case study, we find that OpenAI’s",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 141,
+ 365,
+ 469,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 365,
+ 469,
+ 378
+ ],
+ "score": 1.0,
+ "content": "Codex errs predictably based on how the input prompt is framed, adjusts outputs",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 376,
+ 470,
+ 388
+ ],
+ "score": 1.0,
+ "content": "towards anchors, and is biased towards outputs that mimic frequent training exam-",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 141,
+ 386,
+ 470,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 386,
+ 470,
+ 399
+ ],
+ "score": 1.0,
+ "content": "ples. We then use our framework to elicit high-impact errors such as incorrectly",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 141,
+ 397,
+ 469,
+ 410
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 397,
+ 469,
+ 410
+ ],
+ "score": 1.0,
+ "content": "deleting files. Our results indicate that experimental methodology from cognitive",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 141,
+ 409,
+ 423,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 409,
+ 423,
+ 421
+ ],
+ "score": 1.0,
+ "content": "science can help characterize how machine learning systems behave.1",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 15.5,
+ "bbox_fs": [
+ 140,
+ 267,
+ 471,
+ 421
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 441,
+ 191,
+ 455
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 192,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 192,
+ 457
+ ],
+ "score": 1.0,
+ "content": "1 Introduction",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 467,
+ 504,
+ 501
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 466,
+ 506,
+ 479
+ ],
+ "score": 1.0,
+ "content": "Recent large language models have achieved new, exciting capabilities. In contrast to traditional",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 490
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 478,
+ 506,
+ 490
+ ],
+ "score": 1.0,
+ "content": "classifiers, these models can generate open-ended text, enabling use cases like summarization",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 108,
+ 488,
+ 485,
+ 501
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 488,
+ 485,
+ 501
+ ],
+ "score": 1.0,
+ "content": "[Stiennon et al., 2020], dialog [Thoppilan et al., 2022], and code generation [Chen et al., 2021]",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 466,
+ 506,
+ 501
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 506,
+ 505,
+ 582
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 506,
+ 518
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 506,
+ 506,
+ 518
+ ],
+ "score": 1.0,
+ "content": "The open-ended power of these systems, however, poses new reliability challenges. We must",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 528
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 517,
+ 505,
+ 528
+ ],
+ "score": 1.0,
+ "content": "understand not only when systems err, but also the kinds of errors they make, as some errors are much",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 540
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 527,
+ 505,
+ 540
+ ],
+ "score": 1.0,
+ "content": "more costly than others. For example, erroneous code that does not compile is less dangerous than",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 538,
+ 506,
+ 551
+ ],
+ "score": 1.0,
+ "content": "code that deletes all files in the home directory. Studying how frequently an error occurs is difficult,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 549,
+ 506,
+ 561
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 549,
+ 506,
+ 561
+ ],
+ "score": 1.0,
+ "content": "as the same error (e.g. delete all files) can appear in a wide range of syntactically diverse outputs.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 573
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 559,
+ 505,
+ 573
+ ],
+ "score": 1.0,
+ "content": "In order to better reason about how complex systems err, we need methods to test whether systems",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 571,
+ 495,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 571,
+ 495,
+ 584
+ ],
+ "score": 1.0,
+ "content": "make the same qualitative error across different prompts, even when the generated outputs differ.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 506,
+ 506,
+ 584
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 587,
+ 505,
+ 642
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 600
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 505,
+ 600
+ ],
+ "score": 1.0,
+ "content": "To study these reliability challenges, we primarily focus on code generation models. Such models",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 506,
+ 610
+ ],
+ "score": 1.0,
+ "content": "complete programs from comments, descriptions of code functionality, or initial lines of code.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 622
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 609,
+ 505,
+ 622
+ ],
+ "score": 1.0,
+ "content": "Code generation is particularly amenable to study since it is objective: generated solutions are",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 619,
+ 505,
+ 632
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 619,
+ 505,
+ 632
+ ],
+ "score": 1.0,
+ "content": "unambiguously correct or incorrect. Yet it is also open-ended: the set of programs a model could",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 631,
+ 506,
+ 643
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 631,
+ 506,
+ 643
+ ],
+ "score": 1.0,
+ "content": "output is arbitrarily large, so the rate at which a specific program is outputted is not very descriptive.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 587,
+ 506,
+ 643
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 647,
+ 505,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 505,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 647,
+ 505,
+ 659
+ ],
+ "score": 1.0,
+ "content": "Many of the reliability challenges posed by code generation models, and open-ended systems broadly,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 658,
+ 506,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 506,
+ 669
+ ],
+ "score": 1.0,
+ "content": "also arise when studying qualitative failures in human decision making. These failures, called",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 106,
+ 669,
+ 506,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 669,
+ 506,
+ 682
+ ],
+ "score": 1.0,
+ "content": "cognitive biases, are systematic ways in which humans deviate from rational judgment [Tversky",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 680,
+ 506,
+ 693
+ ],
+ "score": 1.0,
+ "content": "and Kahneman, 1974]. For example, Tversky and Kahneman find that humans inadequately adjust",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 241
+ ],
+ "score": 1.0,
+ "content": "estimates away from initial values, and disproportionately recall distinctive examples. To uncover",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "cognitive biases, Tversky and Kahneman ask questions that are crafted to systematically reveal some",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 251,
+ 507,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 251,
+ 507,
+ 264
+ ],
+ "score": 1.0,
+ "content": "qualitative irrationality. They uncover insights into human behavior from the diverse responses,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 262,
+ 417,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 417,
+ 273
+ ],
+ "score": 1.0,
+ "content": "without complete mechanistic insight into the minds that they aim to analyze.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 105,
+ 647,
+ 506,
+ 693
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "score": 0.953,
+ "type": "image",
+ "image_path": "b810734399a6a31fa69f1d979a6e552103d7ab86daa2ba2d29853274d717cdaf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 99.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 99.66666666666667,
+ 503,
+ 127.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 127.33333333333334,
+ 503,
+ 155.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 160,
+ 505,
+ 204
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 160,
+ 506,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 160,
+ 506,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Illustration of our experimental framework. We use a cognitive bias (framing effect)",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "to inspire a potential code generation failure mode (relying on irrelevant information). We then",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 183,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 183,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "transform inputs in a way that we suspect will elicit the failure mode (prepending sum). We evaluate",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 193,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 193,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "whether the modifications lower accuracy, and if the output is an instance of the targeted failure mode.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 228,
+ 505,
+ 273
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 241
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 241
+ ],
+ "score": 1.0,
+ "content": "estimates away from initial values, and disproportionately recall distinctive examples. To uncover",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 252
+ ],
+ "score": 1.0,
+ "content": "cognitive biases, Tversky and Kahneman ask questions that are crafted to systematically reveal some",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 251,
+ 507,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 251,
+ 507,
+ 264
+ ],
+ "score": 1.0,
+ "content": "qualitative irrationality. They uncover insights into human behavior from the diverse responses,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 262,
+ 417,
+ 273
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 262,
+ 417,
+ 273
+ ],
+ "score": 1.0,
+ "content": "without complete mechanistic insight into the minds that they aim to analyze.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 8.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 277,
+ 505,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "In this work, we extend Tversky and Kahneman’s experimental methodology and results to elicit",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": "failure modes of large code and language models, without relying on complete mechanistic insight",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "into their behavior (Figure 1). Given a potential failure mode (e.g. relying on irrelevant information",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 323
+ ],
+ "score": 1.0,
+ "content": "in the input), we construct a transformation over inputs that largely preserves semantics, but that we",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "suspect will elicit the failure (e.g. prepending an irrelevant function). We first test if the model is",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "sensitive to the transformation, by measuring if it decreases accuracy. Then, we check that the model",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 504,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 504,
+ 356
+ ],
+ "score": 1.0,
+ "content": "outputs have elements that are indicative of the targeted failure (e.g. copies the irrelevant function).",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 359,
+ 506,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "We draw on four different cognitive biases to hypothesize potential failures of OpenAI’s Codex [Chen",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 507,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 130,
+ 384
+ ],
+ "score": 1.0,
+ "content": "et al.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 370,
+ 156,
+ 383
+ ],
+ "score": 0.47,
+ "content": "\\boxed { 2 0 2 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 370,
+ 268,
+ 384
+ ],
+ "score": 1.0,
+ "content": "and Salesforce’s CodeGen",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 370,
+ 360,
+ 383
+ ],
+ "score": 0.47,
+ "content": "[ [ \\mathrm { N i j k a m p ~ e t ~ a l . } ] [ \\overline { { 2 0 2 2 } } ] ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 370,
+ 507,
+ 384
+ ],
+ "score": 1.0,
+ "content": ", then apply our framework to each.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "Our results indicate that these models often rely on irrelevant information when generating solutions,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "adjust solutions towards related-but-incorrect solutions, are biased based on training-set frequencies,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "and reverts to computationally simpler problems when faced with a complex calculation. We also",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "apply our framework to OpenAI’s GPT-3 [Brown et al., 2020], and show that it updates its predictions",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 449,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 449,
+ 439
+ ],
+ "score": 1.0,
+ "content": "towards anchors, and predictably adjusts its responses based on the question framing.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 442,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "Finally, we show that our framework can uncover high-impact errors: errors that are harmful and",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 506,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 506,
+ 465
+ ],
+ "score": 1.0,
+ "content": "difficult to undo. Specifically, we use our framework to systematically generate prompts where Codex",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "erroneously deletes files. Our results indicate that experimental methodology from cognitive science",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 389,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 389,
+ 488
+ ],
+ "score": 1.0,
+ "content": "can help uncover failure modes of complex machine learning systems.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 504,
+ 197,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 198,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 198,
+ 519
+ ],
+ "score": 1.0,
+ "content": "2 Related Work",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 531,
+ 506,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 530,
+ 506,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 530,
+ 506,
+ 542
+ ],
+ "score": 1.0,
+ "content": "Large language models. Recent work has developed large, capable, autoregressive language models,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 554
+ ],
+ "score": 1.0,
+ "content": "which predict future tokens from past tokens [Radford et al., 2019, Wang and Komatsuzaki, 2021,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Brown et al., 2020, Chen et al., 2021, Rae et al., 2021]. These models can be used for open-ended",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "generation tasks such as summarization [Stiennon et al., 2020, Ziegler et al., 2019, Rothe et al., 2020],",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 504,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 504,
+ 587
+ ],
+ "score": 1.0,
+ "content": "dialogue [Ram et al., 2018, Thoppilan et al., 2022], and long form question answering [Fan et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 131,
+ 598
+ ],
+ "score": 1.0,
+ "content": "2019]",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 585,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", among others. Model-generated code has been used to solve both programming and statistics",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 294,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 147,
+ 608
+ ],
+ "score": 1.0,
+ "content": "questions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 596,
+ 294,
+ 609
+ ],
+ "score": 1.0,
+ "content": "[Chen et al., 2021, Tang et al., 2021].",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 612,
+ 505,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": "There is some existing work studying failures of large language models. Benchmarks that measure",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 504,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 504,
+ 636
+ ],
+ "score": 1.0,
+ "content": "model performance on multiple choice questions [Wang et al., 2019b,a, Hendrycks et al., 2021b],",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 634,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 634,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "mathematics [Hendrycks et al., 2021c, Cobbe et al., 2021], long-form question answering [Lin et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 659
+ ],
+ "score": 1.0,
+ "content": "2021, Gabriel et al., 2021, Shuster et al., 2021, Krishna et al., 2021], and coding problems [Hendrycks",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 507,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 128,
+ 667
+ ],
+ "score": 1.0,
+ "content": "et al.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 654,
+ 507,
+ 669
+ ],
+ "score": 1.0,
+ "content": "2021a, Chen et al., 2021] reveal inputs that the model errs on, but not the kind of error it makes.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "Another line of work shows that test-based language models can internalize bias and stereotypes",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 108,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "[Sheng et al., 2019, Nadeem et al., 2020, Groenwold et al., 2020, Blodgett et al., 2021, Gehman",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "et al., 2020], and proposes applying fairness measurements from cognitive social sciences to machine",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "learning systems [Jacobs and Wallach, 2021]. Some work adversarially prompts models to leak",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "training data [Carlini et al., 2020], or output specific content [Wallace et al., 2019, Carlini et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 41.5
+ }
+ ],
+ "page_idx": 1,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "score": 1.0,
+ "content": "2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 155
+ ],
+ "score": 0.953,
+ "type": "image",
+ "image_path": "b810734399a6a31fa69f1d979a6e552103d7ab86daa2ba2d29853274d717cdaf.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 72,
+ 503,
+ 99.66666666666667
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 99.66666666666667,
+ 503,
+ 127.33333333333334
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 127.33333333333334,
+ 503,
+ 155.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 160,
+ 505,
+ 204
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 160,
+ 506,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 160,
+ 506,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Figure 1: Illustration of our experimental framework. We use a cognitive bias (framing effect)",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 172,
+ 505,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 505,
+ 183
+ ],
+ "score": 1.0,
+ "content": "to inspire a potential code generation failure mode (relying on irrelevant information). We then",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 183,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 183,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "transform inputs in a way that we suspect will elicit the failure mode (prepending sum). We evaluate",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 193,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 193,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "whether the modifications lower accuracy, and if the output is an instance of the targeted failure mode.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ }
+ ],
+ "index": 2.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 228,
+ 505,
+ 273
+ ],
+ "lines": [],
+ "index": 8.5,
+ "bbox_fs": [
+ 105,
+ 228,
+ 507,
+ 273
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 277,
+ 505,
+ 355
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 506,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 277,
+ 506,
+ 290
+ ],
+ "score": 1.0,
+ "content": "In this work, we extend Tversky and Kahneman’s experimental methodology and results to elicit",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": "failure modes of large code and language models, without relying on complete mechanistic insight",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 300,
+ 505,
+ 313
+ ],
+ "score": 1.0,
+ "content": "into their behavior (Figure 1). Given a potential failure mode (e.g. relying on irrelevant information",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 323
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 310,
+ 506,
+ 323
+ ],
+ "score": 1.0,
+ "content": "in the input), we construct a transformation over inputs that largely preserves semantics, but that we",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 321,
+ 506,
+ 333
+ ],
+ "score": 1.0,
+ "content": "suspect will elicit the failure (e.g. prepending an irrelevant function). We first test if the model is",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 333,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "sensitive to the transformation, by measuring if it decreases accuracy. Then, we check that the model",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 344,
+ 504,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 344,
+ 504,
+ 356
+ ],
+ "score": 1.0,
+ "content": "outputs have elements that are indicative of the targeted failure (e.g. copies the irrelevant function).",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 105,
+ 277,
+ 506,
+ 356
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 359,
+ 506,
+ 437
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 360,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "We draw on four different cognitive biases to hypothesize potential failures of OpenAI’s Codex [Chen",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 370,
+ 507,
+ 384
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 370,
+ 130,
+ 384
+ ],
+ "score": 1.0,
+ "content": "et al.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 130,
+ 370,
+ 156,
+ 383
+ ],
+ "score": 0.47,
+ "content": "\\boxed { 2 0 2 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 370,
+ 268,
+ 384
+ ],
+ "score": 1.0,
+ "content": "and Salesforce’s CodeGen",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 269,
+ 370,
+ 360,
+ 383
+ ],
+ "score": 0.47,
+ "content": "[ [ \\mathrm { N i j k a m p ~ e t ~ a l . } ] [ \\overline { { 2 0 2 2 } } ] ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 360,
+ 370,
+ 507,
+ 384
+ ],
+ "score": 1.0,
+ "content": ", then apply our framework to each.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 506,
+ 394
+ ],
+ "score": 1.0,
+ "content": "Our results indicate that these models often rely on irrelevant information when generating solutions,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 391,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 391,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "adjust solutions towards related-but-incorrect solutions, are biased based on training-set frequencies,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 403,
+ 505,
+ 415
+ ],
+ "score": 1.0,
+ "content": "and reverts to computationally simpler problems when faced with a complex calculation. We also",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "apply our framework to OpenAI’s GPT-3 [Brown et al., 2020], and show that it updates its predictions",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 424,
+ 449,
+ 439
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 424,
+ 449,
+ 439
+ ],
+ "score": 1.0,
+ "content": "towards anchors, and predictably adjusts its responses based on the question framing.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 360,
+ 507,
+ 439
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 442,
+ 505,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 455
+ ],
+ "score": 1.0,
+ "content": "Finally, we show that our framework can uncover high-impact errors: errors that are harmful and",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 452,
+ 506,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 452,
+ 506,
+ 465
+ ],
+ "score": 1.0,
+ "content": "difficult to undo. Specifically, we use our framework to systematically generate prompts where Codex",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 463,
+ 505,
+ 477
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 463,
+ 505,
+ 477
+ ],
+ "score": 1.0,
+ "content": "erroneously deletes files. Our results indicate that experimental methodology from cognitive science",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 474,
+ 389,
+ 488
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 389,
+ 488
+ ],
+ "score": 1.0,
+ "content": "can help uncover failure modes of complex machine learning systems.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 442,
+ 506,
+ 488
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 504,
+ 197,
+ 517
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 198,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 198,
+ 519
+ ],
+ "score": 1.0,
+ "content": "2 Related Work",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 531,
+ 506,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 530,
+ 506,
+ 542
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 530,
+ 506,
+ 542
+ ],
+ "score": 1.0,
+ "content": "Large language models. Recent work has developed large, capable, autoregressive language models,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 554
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 504,
+ 554
+ ],
+ "score": 1.0,
+ "content": "which predict future tokens from past tokens [Radford et al., 2019, Wang and Komatsuzaki, 2021,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 552,
+ 505,
+ 565
+ ],
+ "score": 1.0,
+ "content": "Brown et al., 2020, Chen et al., 2021, Rae et al., 2021]. These models can be used for open-ended",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 104,
+ 563,
+ 505,
+ 576
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 563,
+ 505,
+ 576
+ ],
+ "score": 1.0,
+ "content": "generation tasks such as summarization [Stiennon et al., 2020, Ziegler et al., 2019, Rothe et al., 2020],",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 573,
+ 504,
+ 587
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 573,
+ 504,
+ 587
+ ],
+ "score": 1.0,
+ "content": "dialogue [Ram et al., 2018, Thoppilan et al., 2022], and long form question answering [Fan et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 598
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 131,
+ 598
+ ],
+ "score": 1.0,
+ "content": "2019]",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 132,
+ 585,
+ 506,
+ 597
+ ],
+ "score": 1.0,
+ "content": ", among others. Model-generated code has been used to solve both programming and statistics",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 596,
+ 294,
+ 609
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 596,
+ 147,
+ 608
+ ],
+ "score": 1.0,
+ "content": "questions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 596,
+ 294,
+ 609
+ ],
+ "score": 1.0,
+ "content": "[Chen et al., 2021, Tang et al., 2021].",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 104,
+ 530,
+ 506,
+ 609
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 612,
+ 505,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": "There is some existing work studying failures of large language models. Benchmarks that measure",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 504,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 504,
+ 636
+ ],
+ "score": 1.0,
+ "content": "model performance on multiple choice questions [Wang et al., 2019b,a, Hendrycks et al., 2021b],",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 634,
+ 505,
+ 647
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 634,
+ 505,
+ 647
+ ],
+ "score": 1.0,
+ "content": "mathematics [Hendrycks et al., 2021c, Cobbe et al., 2021], long-form question answering [Lin et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 659
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 659
+ ],
+ "score": 1.0,
+ "content": "2021, Gabriel et al., 2021, Shuster et al., 2021, Krishna et al., 2021], and coding problems [Hendrycks",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 654,
+ 507,
+ 669
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 656,
+ 128,
+ 667
+ ],
+ "score": 1.0,
+ "content": "et al.,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 129,
+ 654,
+ 507,
+ 669
+ ],
+ "score": 1.0,
+ "content": "2021a, Chen et al., 2021] reveal inputs that the model errs on, but not the kind of error it makes.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 679
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 666,
+ 506,
+ 679
+ ],
+ "score": 1.0,
+ "content": "Another line of work shows that test-based language models can internalize bias and stereotypes",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 108,
+ 677,
+ 506,
+ 690
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 677,
+ 506,
+ 690
+ ],
+ "score": 1.0,
+ "content": "[Sheng et al., 2019, Nadeem et al., 2020, Groenwold et al., 2020, Blodgett et al., 2021, Gehman",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 689,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "et al., 2020], and proposes applying fairness measurements from cognitive social sciences to machine",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 711
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 711
+ ],
+ "score": 1.0,
+ "content": "learning systems [Jacobs and Wallach, 2021]. Some work adversarially prompts models to leak",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 505,
+ 723
+ ],
+ "score": 1.0,
+ "content": "training data [Carlini et al., 2020], or output specific content [Wallace et al., 2019, Carlini et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 131,
+ 295
+ ],
+ "score": 0.35,
+ "content": "\\boxed { 2 0 2 0 } ]",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 131,
+ 281,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": ". And a final line of work identifies additional potential failures of current and future machine",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 453,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 453,
+ 306
+ ],
+ "score": 1.0,
+ "content": "learning systems [Bender et al., 2021, Bommasani et al., 2021, Weidinger et al., 2021].",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 41.5,
+ "bbox_fs": [
+ 105,
+ 611,
+ 507,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "score": 0.68,
+ "type": "image",
+ "image_path": "b0ac32b01cb78509c5227342d766cad25cb336596ed9fdc4bd31985e4f6b5b89.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 102.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 102.0,
+ 503,
+ 135.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 135.0,
+ 503,
+ 168.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 172,
+ 506,
+ 262
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 404,
+ 186
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Left. Example of a HumanEval problem from Chen et al. [2021]",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 172,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": ". The problem contains a",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 183,
+ 506,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 183,
+ 506,
+ 197
+ ],
+ "score": 1.0,
+ "content": "prompt (blue), a canonical solution to the prompt (green), and a few test-cases (black). The prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 507,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 507,
+ 207
+ ],
+ "score": 1.0,
+ "content": "contains two components: a function signature (first line), and a docstring (remaining lines). Right.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 206,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 206,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Illustration of our framing experiment. The transformed prompt (everything above the black line)",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 217,
+ 505,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 505,
+ 229
+ ],
+ "score": 1.0,
+ "content": "contains an irrelevant preceding function (IPF) prepended to a prompt from HumanEval (blue). The",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "IPF contains a randomly chosen prompt from HumanEval (purple) and a framing line (red). The",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "output Codex generates (below the black line) matches the framing line. When we omit the random",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 249,
+ 493,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 493,
+ 262
+ ],
+ "score": 1.0,
+ "content": "HumanEval prompt and the framing line (leaving only blue), Codex produces the correct output.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 282,
+ 504,
+ 306
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 131,
+ 295
+ ],
+ "score": 0.35,
+ "content": "\\boxed { 2 0 2 0 } ]",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 281,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": ". And a final line of work identifies additional potential failures of current and future machine",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 291,
+ 453,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 291,
+ 453,
+ 306
+ ],
+ "score": 1.0,
+ "content": "learning systems [Bender et al., 2021, Bommasani et al., 2021, Weidinger et al., 2021].",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 308,
+ 505,
+ 366
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "Cognitive biases. Tversky and Kahneman [1974] define human cognitive biases: systematic patterns",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "of deviation from rational judgment. They observe that humans employ heuristics when computing",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "probabilities or assessing values, and that these heuristics lead to predictable errors. Follow-up work",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 356
+ ],
+ "score": 1.0,
+ "content": "has added to, refined, and validated the set of known cognitive biases [Tversky and Kahneman, 1973,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 351,
+ 492,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 351,
+ 492,
+ 367
+ ],
+ "score": 1.0,
+ "content": "1981, Strack et al., 1988, Kahneman and Frederick, 2002, Windhager et al., 2010, Meyer, 2014].",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 369,
+ 505,
+ 435
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 504,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 476,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Some known failure modes of large language models resemble cognitive biases. Zhao et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 370,
+ 504,
+ 382
+ ],
+ "score": 1.0,
+ "content": "[2021]",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 163,
+ 392
+ ],
+ "score": 1.0,
+ "content": "and Liu et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 379,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "[2021] show that the specific random samples used for few-shot learning can change",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 390,
+ 505,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 435,
+ 403
+ ],
+ "score": 1.0,
+ "content": "GPT-3’s prediction on binary and multiple choice tasks. Similarly, Wallace et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 437,
+ 392,
+ 505,
+ 403
+ ],
+ "score": 1.0,
+ "content": "[2019] show that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "innocuous prompts can routinely generate toxic model output. Our framework builds on this work by",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "(i) identifying the link to cognitive biases, (ii) focusing on open-ended generation, and (iii) leveraging",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 424,
+ 456,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 456,
+ 436
+ ],
+ "score": 1.0,
+ "content": "Tversky and Kahneman’s experimental methodology to elicit qualitative failure modes.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 450,
+ 282,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 448,
+ 283,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 448,
+ 283,
+ 468
+ ],
+ "score": 1.0,
+ "content": "3 Code Generation Experiments",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 474,
+ 162,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 162,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 162,
+ 487
+ ],
+ "score": 1.0,
+ "content": "3.1 Models",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 494,
+ 505,
+ 539
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 507,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 507,
+ 508
+ ],
+ "score": 1.0,
+ "content": "We study two code models: OpenAI’s Codex [Chen et al., 2021], and Salesforce’s CodeGen. [Nijkamp",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 507,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 507,
+ 519
+ ],
+ "score": 1.0,
+ "content": "et al., 2022]. Both models are autoregressive—given a sequence of previous tokens, they predict",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "the next token. Practitioners query these code models with partial programs, docstrings, or function",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 528,
+ 289,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 289,
+ 541
+ ],
+ "score": 1.0,
+ "content": "signatures, and obtain completions as output.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 543,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "score": 1.0,
+ "content": "Codex. We study OpenAI’s Codex, a large language model trained to generate code from docstrings",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 107,
+ 555,
+ 506,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 555,
+ 506,
+ 568
+ ],
+ "score": 1.0,
+ "content": "[Chen et al., 2021]. We use the OpenAI API to query the “davinci-001” version of Codex, and use",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 579
+ ],
+ "score": 1.0,
+ "content": "greedy decoding to generate solutions. Details of this model architecture are not public, but it is likely",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "similar to the largest model from Chen et al. [2021]: a 12B parameter version of GPT-3 [Brown et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 369,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 369,
+ 599
+ ],
+ "score": 1.0,
+ "content": "2020] that is fine-tuned on GitHub instead of the CommonCrawl.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 638
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 505,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 505,
+ 616
+ ],
+ "score": 1.0,
+ "content": "CodeGen. We additionally study the 6.2 billion parameter “mono” version of CodeGen, which is",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "trained on text data and fine-tuned on GitHub. Unlike Codex, the weights of CodeGen are publicly",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 626,
+ 451,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 626,
+ 451,
+ 639
+ ],
+ "score": 1.0,
+ "content": "available,2 so we run inference locally. We use greedy decoding to generate solutions.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 650,
+ 184,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 185,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 185,
+ 663
+ ],
+ "score": 1.0,
+ "content": "3.2 Benchmarks",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 670,
+ 505,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 506,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 506,
+ 683
+ ],
+ "score": 1.0,
+ "content": "In order to identify whether code models some failure mode, we need to generate prompts that elicit",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 680,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 680,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": "that failure. To do so, we systematically apply transformations to standard prompts. We use two",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 691,
+ 436,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 436,
+ 705
+ ],
+ "score": 1.0,
+ "content": "benchmarks as sources of prompts to transform: HumanEval, and MathEquations.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "page_idx": 2,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 122,
+ 712,
+ 298,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 121,
+ 711,
+ 299,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 121,
+ 711,
+ 299,
+ 723
+ ],
+ "score": 1.0,
+ "content": "2 https://github.com/salesforce/CodeGen",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "3",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 168
+ ],
+ "score": 0.68,
+ "type": "image",
+ "image_path": "b0ac32b01cb78509c5227342d766cad25cb336596ed9fdc4bd31985e4f6b5b89.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 69,
+ 503,
+ 102.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 102.0,
+ 503,
+ 135.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 135.0,
+ 503,
+ 168.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 172,
+ 506,
+ 262
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 506,
+ 186
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 172,
+ 404,
+ 186
+ ],
+ "score": 1.0,
+ "content": "Figure 2: Left. Example of a HumanEval problem from Chen et al. [2021]",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 405,
+ 172,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": ". The problem contains a",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 183,
+ 506,
+ 197
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 183,
+ 506,
+ 197
+ ],
+ "score": 1.0,
+ "content": "prompt (blue), a canonical solution to the prompt (green), and a few test-cases (black). The prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 195,
+ 507,
+ 207
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 195,
+ 507,
+ 207
+ ],
+ "score": 1.0,
+ "content": "contains two components: a function signature (first line), and a docstring (remaining lines). Right.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 206,
+ 506,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 206,
+ 506,
+ 218
+ ],
+ "score": 1.0,
+ "content": "Illustration of our framing experiment. The transformed prompt (everything above the black line)",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 217,
+ 505,
+ 229
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 217,
+ 505,
+ 229
+ ],
+ "score": 1.0,
+ "content": "contains an irrelevant preceding function (IPF) prepended to a prompt from HumanEval (blue). The",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 240
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 505,
+ 240
+ ],
+ "score": 1.0,
+ "content": "IPF contains a randomly chosen prompt from HumanEval (purple) and a framing line (red). The",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 239,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "output Codex generates (below the black line) matches the framing line. When we omit the random",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 249,
+ 493,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 249,
+ 493,
+ 262
+ ],
+ "score": 1.0,
+ "content": "HumanEval prompt and the framing line (leaving only blue), Codex produces the correct output.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 282,
+ 504,
+ 306
+ ],
+ "lines": [],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 281,
+ 505,
+ 306
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 308,
+ 505,
+ 366
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 308,
+ 506,
+ 322
+ ],
+ "score": 1.0,
+ "content": "Cognitive biases. Tversky and Kahneman [1974] define human cognitive biases: systematic patterns",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 318,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "of deviation from rational judgment. They observe that humans employ heuristics when computing",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 343
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 330,
+ 506,
+ 343
+ ],
+ "score": 1.0,
+ "content": "probabilities or assessing values, and that these heuristics lead to predictable errors. Follow-up work",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 340,
+ 506,
+ 356
+ ],
+ "score": 1.0,
+ "content": "has added to, refined, and validated the set of known cognitive biases [Tversky and Kahneman, 1973,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 104,
+ 351,
+ 492,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 351,
+ 492,
+ 367
+ ],
+ "score": 1.0,
+ "content": "1981, Strack et al., 1988, Kahneman and Frederick, 2002, Windhager et al., 2010, Meyer, 2014].",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 104,
+ 308,
+ 506,
+ 367
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 369,
+ 505,
+ 435
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 504,
+ 382
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 368,
+ 476,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Some known failure modes of large language models resemble cognitive biases. Zhao et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 478,
+ 370,
+ 504,
+ 382
+ ],
+ "score": 1.0,
+ "content": "[2021]",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 379,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 163,
+ 392
+ ],
+ "score": 1.0,
+ "content": "and Liu et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 164,
+ 379,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "[2021] show that the specific random samples used for few-shot learning can change",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 390,
+ 505,
+ 403
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 390,
+ 435,
+ 403
+ ],
+ "score": 1.0,
+ "content": "GPT-3’s prediction on binary and multiple choice tasks. Similarly, Wallace et al.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 437,
+ 392,
+ 505,
+ 403
+ ],
+ "score": 1.0,
+ "content": "[2019] show that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 414
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 401,
+ 505,
+ 414
+ ],
+ "score": 1.0,
+ "content": "innocuous prompts can routinely generate toxic model output. Our framework builds on this work by",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 426
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 412,
+ 505,
+ 426
+ ],
+ "score": 1.0,
+ "content": "(i) identifying the link to cognitive biases, (ii) focusing on open-ended generation, and (iii) leveraging",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 424,
+ 456,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 456,
+ 436
+ ],
+ "score": 1.0,
+ "content": "Tversky and Kahneman’s experimental methodology to elicit qualitative failure modes.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 368,
+ 505,
+ 436
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 450,
+ 282,
+ 464
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 448,
+ 283,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 448,
+ 283,
+ 468
+ ],
+ "score": 1.0,
+ "content": "3 Code Generation Experiments",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 106,
+ 474,
+ 162,
+ 486
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 162,
+ 487
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 474,
+ 162,
+ 487
+ ],
+ "score": 1.0,
+ "content": "3.1 Models",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 494,
+ 505,
+ 539
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 507,
+ 508
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 494,
+ 507,
+ 508
+ ],
+ "score": 1.0,
+ "content": "We study two code models: OpenAI’s Codex [Chen et al., 2021], and Salesforce’s CodeGen. [Nijkamp",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 507,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 507,
+ 519
+ ],
+ "score": 1.0,
+ "content": "et al., 2022]. Both models are autoregressive—given a sequence of previous tokens, they predict",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 517,
+ 505,
+ 529
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 517,
+ 505,
+ 529
+ ],
+ "score": 1.0,
+ "content": "the next token. Practitioners query these code models with partial programs, docstrings, or function",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 528,
+ 289,
+ 541
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 528,
+ 289,
+ 541
+ ],
+ "score": 1.0,
+ "content": "signatures, and obtain completions as output.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 494,
+ 507,
+ 541
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 543,
+ 505,
+ 599
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 543,
+ 506,
+ 557
+ ],
+ "score": 1.0,
+ "content": "Codex. We study OpenAI’s Codex, a large language model trained to generate code from docstrings",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 107,
+ 555,
+ 506,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 555,
+ 506,
+ 568
+ ],
+ "score": 1.0,
+ "content": "[Chen et al., 2021]. We use the OpenAI API to query the “davinci-001” version of Codex, and use",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 579
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 565,
+ 506,
+ 579
+ ],
+ "score": 1.0,
+ "content": "greedy decoding to generate solutions. Details of this model architecture are not public, but it is likely",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 576,
+ 505,
+ 590
+ ],
+ "score": 1.0,
+ "content": "similar to the largest model from Chen et al. [2021]: a 12B parameter version of GPT-3 [Brown et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 587,
+ 369,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 587,
+ 369,
+ 599
+ ],
+ "score": 1.0,
+ "content": "2020] that is fine-tuned on GitHub instead of the CommonCrawl.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 543,
+ 506,
+ 599
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 604,
+ 505,
+ 638
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 505,
+ 616
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 604,
+ 505,
+ 616
+ ],
+ "score": 1.0,
+ "content": "CodeGen. We additionally study the 6.2 billion parameter “mono” version of CodeGen, which is",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 627
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 505,
+ 627
+ ],
+ "score": 1.0,
+ "content": "trained on text data and fine-tuned on GitHub. Unlike Codex, the weights of CodeGen are publicly",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 626,
+ 451,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 626,
+ 451,
+ 639
+ ],
+ "score": 1.0,
+ "content": "available,2 so we run inference locally. We use greedy decoding to generate solutions.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 36,
+ "bbox_fs": [
+ 105,
+ 604,
+ 505,
+ 639
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 650,
+ 184,
+ 662
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 185,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 649,
+ 185,
+ 663
+ ],
+ "score": 1.0,
+ "content": "3.2 Benchmarks",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 670,
+ 505,
+ 703
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 506,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 670,
+ 506,
+ 683
+ ],
+ "score": 1.0,
+ "content": "In order to identify whether code models some failure mode, we need to generate prompts that elicit",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 105,
+ 680,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 680,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": "that failure. To do so, we systematically apply transformations to standard prompts. We use two",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 691,
+ 436,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 691,
+ 436,
+ 705
+ ],
+ "score": 1.0,
+ "content": "benchmarks as sources of prompts to transform: HumanEval, and MathEquations.",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40,
+ "bbox_fs": [
+ 105,
+ 670,
+ 506,
+ 705
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "score": 0.978,
+ "html": "
| Framing Line | Model | ORIGINAL | FRAMED | ORIGINAL | FRAMED |
| raise NotImplemented | CODEX | 32.9 | 2.4 | 1.4 | 91.7 |
| CODEGEN | 25.6 | 1.5 | 0.0 | 79.3 |
| pass | CoDEX | 32.9 | 3.0 | 9.7 | 92.7 |
| CODEGEN | 25.6 | 2.1 | 0.0 | 78.7 |
| assert False | CODEX | 32.9 | 3.3 | 0.0 | 92.7 |
| CODEGEN | 25.6 | 4.2 | 0.1 | 72.6 |
| return False | CODEX | 32.9 | 4.9 | 11.5 | 65.6 |
| CODEGEN | 25.6 | 3.6 | 0.0 | 64.6 |
| print("Hello world!") | CODEX | 32.9 | 10.6 | 0.0 | 62.2 |
| CODEGEN | 25.6 | 11.0 | 0.0 | 58.2 |
",
+ "type": "table",
+ "image_path": "a950596f8bca75dcae5dc344e398135cd6410d2b7236ca7ff919a367c6926c0f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 131.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 131.33333333333334,
+ 502,
+ 183.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 183.66666666666669,
+ 502,
+ 236.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_footnote",
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 293
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "Table 1: Results of the framing experiments. We compare functional accuracy and the rate at which",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "framing line is outputted over HumanEval with (framed) and without (original) irrelevant preceding",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 506,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 506,
+ 274
+ ],
+ "score": 1.0,
+ "content": "functions. We find that the irrelevant preceding functions lower functional accuracy across all framing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "lines for Codex and CodeGen. Moreover, we find that the outputted function often appears verbatim",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 496,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 496,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in the generated output, suggesting that both models rely on irrelevant information in the prompt.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 504,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 504,
+ 335
+ ],
+ "score": 1.0,
+ "content": "HumanEval. We use the HumanEval benchmark as a diverse source of “normal” prompts [Chen et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 130,
+ 345
+ ],
+ "score": 0.35,
+ "content": "\\boxed { 2 0 2 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 333,
+ 505,
+ 346
+ ],
+ "score": 1.0,
+ "content": ". HumanEval contains 164 programming problems, each of which includes a function signature",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "and a docstring. The docstring contains an English description of the desired functionality and a few",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "example input-output pairs. HumanEval also contains a canonical solution for each program, which",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 365,
+ 440,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 177,
+ 379
+ ],
+ "score": 1.0,
+ "content": "we use in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 177,
+ 365,
+ 202,
+ 378
+ ],
+ "score": 0.42,
+ "content": "\\underline { { \\bar { \\vert 3 . 3 . 2 \\vert } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 366,
+ 440,
+ 379
+ ],
+ "score": 1.0,
+ "content": "We give an example problem from HumanEval in Figure 2.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 382,
+ 505,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "MathEquations. We also curate a set of prompts of basic arithmetic functions. For example, we",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "prompt Codex to “Write a function that sums the squares of its inputs”, or “Write a function that",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 485,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 456,
+ 416
+ ],
+ "score": 1.0,
+ "content": "sums its inputs called product_plus_five”. Further details are given in Sections 3.3.3 an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 404,
+ 485,
+ 416
+ ],
+ "score": 1.0,
+ "content": "3.3.4.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 429,
+ 204,
+ 441
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 426,
+ 205,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 426,
+ 205,
+ 444
+ ],
+ "score": 1.0,
+ "content": "3.3 Empirical results",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "In this section, we show how cognitive biases can (i) inspire hypotheses for potential failure modes,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "and (ii) help us design experiments to test these hypotheses. Our approach has three steps. First,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "we construct a transformation over prompts that largely preserves semantics, but that we suspect",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 504,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 504,
+ 494
+ ],
+ "score": 1.0,
+ "content": "will elicit a specific cognitive-bias-inspired failure mode. Next, we measure if code models are",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "sensitive to the transformation, by measuring the decrease in accuracy. And finally, we check that the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "generated output has elements that are indicative of the targeted failure mode. Our approach mirrors",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "the high-level methodology from Tversky and Kahneman [1974]; we empirically elicit specific failure",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 503,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 503,
+ 539
+ ],
+ "score": 1.0,
+ "content": "modes using targeted prompts, without complete mechanistic insight into the system that we study.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 506,
+ 587
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 502,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 502,
+ 555
+ ],
+ "score": 1.0,
+ "content": "We draw inspiration from four cognitive biases: the framing effect (Tversky and Kahneman [1981]",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 551,
+ 506,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 375,
+ 566
+ ],
+ "score": 1.0,
+ "content": "Section 3.3.1), anchoring (Tversky and Kahneman [1974]; Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 552,
+ 401,
+ 565
+ ],
+ "score": 0.51,
+ "content": "\\underline { { \\overline { { | 3 . 3 . 2 ) } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 551,
+ 506,
+ 566
+ ],
+ "score": 1.0,
+ "content": ", the availability heuristic",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 107,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "(Tversky and Kahneman [1973]; Section 3.3.3), and attribute substitution (Kahneman and Frederick",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 574,
+ 196,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 574,
+ 196,
+ 588
+ ],
+ "score": 1.0,
+ "content": "[2002]; Section 3.3.4).",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 599,
+ 255,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 256,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 256,
+ 613
+ ],
+ "score": 1.0,
+ "content": "3.3.1 Inspiration: Framing effect",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 617,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "We first draw inspiration from the framing effect: predictable shifts in human responses when the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "same problem is framed in different ways [Tversky and Kahneman, 1981]. In their study identifying",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 249,
+ 651
+ ],
+ "score": 1.0,
+ "content": "the effect, Tversky and Kahneman",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 640,
+ 505,
+ 652
+ ],
+ "score": 1.0,
+ "content": "[1981] find that subjects favor certainly saving 200 people over",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 219,
+ 663
+ ],
+ "score": 1.0,
+ "content": "saving 600 with probability",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 651,
+ 232,
+ 661
+ ],
+ "score": 0.42,
+ "content": "\\overline { { 1 / 3 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 650,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": ", yet prefer losing 600 with probability 2/3 over certainly losing 400",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 661,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 661,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "(even though these are equivalent). At its core, the framing effect shows how humans can rely on",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 358,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 358,
+ 685
+ ],
+ "score": 1.0,
+ "content": "semantically irrelevant information when they make decisions.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "Using the framing effect as inspiration, we hypothesize that code generation models may generate",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "solutions exclusively from irrelevant information in the prompt. To elicit this failure, we transform",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "HumanEval prompts by prepending irrelevant preceding functions. Specifically, to generate irrelevant",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "page_idx": 3,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 11,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 236
+ ],
+ "score": 0.978,
+ "html": "| Framing Line | Model | ORIGINAL | FRAMED | ORIGINAL | FRAMED |
| raise NotImplemented | CODEX | 32.9 | 2.4 | 1.4 | 91.7 |
| CODEGEN | 25.6 | 1.5 | 0.0 | 79.3 |
| pass | CoDEX | 32.9 | 3.0 | 9.7 | 92.7 |
| CODEGEN | 25.6 | 2.1 | 0.0 | 78.7 |
| assert False | CODEX | 32.9 | 3.3 | 0.0 | 92.7 |
| CODEGEN | 25.6 | 4.2 | 0.1 | 72.6 |
| return False | CODEX | 32.9 | 4.9 | 11.5 | 65.6 |
| CODEGEN | 25.6 | 3.6 | 0.0 | 64.6 |
| print("Hello world!") | CODEX | 32.9 | 10.6 | 0.0 | 62.2 |
| CODEGEN | 25.6 | 11.0 | 0.0 | 58.2 |
",
+ "type": "table",
+ "image_path": "a950596f8bca75dcae5dc344e398135cd6410d2b7236ca7ff919a367c6926c0f.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 79,
+ 502,
+ 131.33333333333334
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 131.33333333333334,
+ 502,
+ 183.66666666666669
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 183.66666666666669,
+ 502,
+ 236.00000000000003
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "table_footnote",
+ "bbox": [
+ 106,
+ 237,
+ 505,
+ 293
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 237,
+ 505,
+ 250
+ ],
+ "score": 1.0,
+ "content": "Table 1: Results of the framing experiments. We compare functional accuracy and the rate at which",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 262
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 247,
+ 505,
+ 262
+ ],
+ "score": 1.0,
+ "content": "framing line is outputted over HumanEval with (framed) and without (original) irrelevant preceding",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 257,
+ 506,
+ 274
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 257,
+ 506,
+ 274
+ ],
+ "score": 1.0,
+ "content": "functions. We find that the irrelevant preceding functions lower functional accuracy across all framing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 269,
+ 505,
+ 283
+ ],
+ "score": 1.0,
+ "content": "lines for Codex and CodeGen. Moreover, we find that the outputted function often appears verbatim",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 281,
+ 496,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 281,
+ 496,
+ 294
+ ],
+ "score": 1.0,
+ "content": "in the generated output, suggesting that both models rely on irrelevant information in the prompt.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 3.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 378
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 504,
+ 335
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 504,
+ 335
+ ],
+ "score": 1.0,
+ "content": "HumanEval. We use the HumanEval benchmark as a diverse source of “normal” prompts [Chen et al.,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 505,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 130,
+ 345
+ ],
+ "score": 0.35,
+ "content": "\\boxed { 2 0 2 1 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 131,
+ 333,
+ 505,
+ 346
+ ],
+ "score": 1.0,
+ "content": ". HumanEval contains 164 programming problems, each of which includes a function signature",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 505,
+ 357
+ ],
+ "score": 1.0,
+ "content": "and a docstring. The docstring contains an English description of the desired functionality and a few",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 354,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "example input-output pairs. HumanEval also contains a canonical solution for each program, which",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 365,
+ 440,
+ 379
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 366,
+ 177,
+ 379
+ ],
+ "score": 1.0,
+ "content": "we use in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 177,
+ 365,
+ 202,
+ 378
+ ],
+ "score": 0.42,
+ "content": "\\underline { { \\bar { \\vert 3 . 3 . 2 \\vert } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 203,
+ 366,
+ 440,
+ 379
+ ],
+ "score": 1.0,
+ "content": "We give an example problem from HumanEval in Figure 2.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 321,
+ 505,
+ 379
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 382,
+ 505,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 382,
+ 505,
+ 394
+ ],
+ "score": 1.0,
+ "content": "MathEquations. We also curate a set of prompts of basic arithmetic functions. For example, we",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 392,
+ 506,
+ 405
+ ],
+ "score": 1.0,
+ "content": "prompt Codex to “Write a function that sums the squares of its inputs”, or “Write a function that",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 485,
+ 416
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 456,
+ 416
+ ],
+ "score": 1.0,
+ "content": "sums its inputs called product_plus_five”. Further details are given in Sections 3.3.3 an",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 459,
+ 404,
+ 485,
+ 416
+ ],
+ "score": 1.0,
+ "content": "3.3.4.",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 14,
+ "bbox_fs": [
+ 105,
+ 382,
+ 506,
+ 416
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 429,
+ 204,
+ 441
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 426,
+ 205,
+ 444
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 426,
+ 205,
+ 444
+ ],
+ "score": 1.0,
+ "content": "3.3 Empirical results",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 538
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 506,
+ 462
+ ],
+ "score": 1.0,
+ "content": "In this section, we show how cognitive biases can (i) inspire hypotheses for potential failure modes,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 461,
+ 506,
+ 473
+ ],
+ "score": 1.0,
+ "content": "and (ii) help us design experiments to test these hypotheses. Our approach has three steps. First,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 105,
+ 472,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 472,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "we construct a transformation over prompts that largely preserves semantics, but that we suspect",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 483,
+ 504,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 483,
+ 504,
+ 494
+ ],
+ "score": 1.0,
+ "content": "will elicit a specific cognitive-bias-inspired failure mode. Next, we measure if code models are",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 505
+ ],
+ "score": 1.0,
+ "content": "sensitive to the transformation, by measuring the decrease in accuracy. And finally, we check that the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 517
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 504,
+ 506,
+ 517
+ ],
+ "score": 1.0,
+ "content": "generated output has elements that are indicative of the targeted failure mode. Our approach mirrors",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 505,
+ 527
+ ],
+ "score": 1.0,
+ "content": "the high-level methodology from Tversky and Kahneman [1974]; we empirically elicit specific failure",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 525,
+ 503,
+ 539
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 525,
+ 503,
+ 539
+ ],
+ "score": 1.0,
+ "content": "modes using targeted prompts, without complete mechanistic insight into the system that we study.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 105,
+ 450,
+ 506,
+ 539
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 542,
+ 506,
+ 587
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 502,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 541,
+ 502,
+ 555
+ ],
+ "score": 1.0,
+ "content": "We draw inspiration from four cognitive biases: the framing effect (Tversky and Kahneman [1981]",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 551,
+ 506,
+ 566
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 551,
+ 375,
+ 566
+ ],
+ "score": 1.0,
+ "content": "Section 3.3.1), anchoring (Tversky and Kahneman [1974]; Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 376,
+ 552,
+ 401,
+ 565
+ ],
+ "score": 0.51,
+ "content": "\\underline { { \\overline { { | 3 . 3 . 2 ) } } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 402,
+ 551,
+ 506,
+ 566
+ ],
+ "score": 1.0,
+ "content": ", the availability heuristic",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 107,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "(Tversky and Kahneman [1973]; Section 3.3.3), and attribute substitution (Kahneman and Frederick",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 107,
+ 574,
+ 196,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 574,
+ 196,
+ 588
+ ],
+ "score": 1.0,
+ "content": "[2002]; Section 3.3.4).",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 26.5,
+ "bbox_fs": [
+ 105,
+ 541,
+ 506,
+ 588
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 599,
+ 255,
+ 611
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 256,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 597,
+ 256,
+ 613
+ ],
+ "score": 1.0,
+ "content": "3.3.1 Inspiration: Framing effect",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 617,
+ 505,
+ 684
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 617,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "We first draw inspiration from the framing effect: predictable shifts in human responses when the",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 627,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "same problem is framed in different ways [Tversky and Kahneman, 1981]. In their study identifying",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 640,
+ 505,
+ 652
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 640,
+ 249,
+ 651
+ ],
+ "score": 1.0,
+ "content": "the effect, Tversky and Kahneman",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 250,
+ 640,
+ 505,
+ 652
+ ],
+ "score": 1.0,
+ "content": "[1981] find that subjects favor certainly saving 200 people over",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 650,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 650,
+ 219,
+ 663
+ ],
+ "score": 1.0,
+ "content": "saving 600 with probability",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 219,
+ 651,
+ 232,
+ 661
+ ],
+ "score": 0.42,
+ "content": "\\overline { { 1 / 3 } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 650,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": ", yet prefer losing 600 with probability 2/3 over certainly losing 400",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 661,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 661,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "(even though these are equivalent). At its core, the framing effect shows how humans can rely on",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 673,
+ 358,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 673,
+ 358,
+ 685
+ ],
+ "score": 1.0,
+ "content": "semantically irrelevant information when they make decisions.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 105,
+ 617,
+ 505,
+ 685
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 689,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 689,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "Using the framing effect as inspiration, we hypothesize that code generation models may generate",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "solutions exclusively from irrelevant information in the prompt. To elicit this failure, we transform",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 505,
+ 724
+ ],
+ "score": 1.0,
+ "content": "HumanEval prompts by prepending irrelevant preceding functions. Specifically, to generate irrelevant",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "preceding functions, we combine a random prompt from HumanEval with a framing line. We test",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 506,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 506,
+ 244
+ ],
+ "score": 1.0,
+ "content": "five framing lines: raise NotImplementedError , pass , assert False , return False , and",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "print(\"Hello world!\") . We first check that prepending these irrelevant preceding functions",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 506,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 506,
+ 267
+ ],
+ "score": 1.0,
+ "content": "decreases functional accuracy.3 Next, to test if models relied on irrelevant information in the prompt,",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 499,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 499,
+ 277
+ ],
+ "score": 1.0,
+ "content": "we measure how much more frequently the framing line appears verbatim in the generated output.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 105,
+ 689,
+ 505,
+ 724
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "score": 0.937,
+ "type": "image",
+ "image_path": "3ca12ceac2dfdc76f5b5cb25925443fec2d6975fface7597f9f0a393592f8cfd.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 88.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 88.0,
+ 502,
+ 107.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 107.0,
+ 502,
+ 126.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 197
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Illustration of our anchoring experiment using a real example (expanded in Figure 7). We",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 155
+ ],
+ "score": 1.0,
+ "content": "construct the anchor function (left) by taking the function signature from the HumanEval prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 153,
+ 506,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 180,
+ 165
+ ],
+ "score": 1.0,
+ "content": "(blue), appending",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 155,
+ 188,
+ 163
+ ],
+ "score": 0.62,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 153,
+ 506,
+ 165
+ ],
+ "score": 1.0,
+ "content": "lines of the canonical solution (green), then adding anchoring lines (red). We",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "construct the full prompt (center) by combining the anchor function, the original HumanEval prompt,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 154,
+ 187
+ ],
+ "score": 1.0,
+ "content": "and the first",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 177,
+ 162,
+ 185
+ ],
+ "score": 0.75,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 176,
+ 505,
+ 187
+ ],
+ "score": 1.0,
+ "content": "lines of the canonical solution. The solution Codex generates (right) combines elements",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 185,
+ 501,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 185,
+ 501,
+ 199
+ ],
+ "score": 1.0,
+ "content": "of a canonical solution (checks condition and adds to ret.), with the anchor function (for var loop).",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 275
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 221,
+ 505,
+ 232
+ ],
+ "score": 1.0,
+ "content": "preceding functions, we combine a random prompt from HumanEval with a framing line. We test",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 230,
+ 506,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 230,
+ 506,
+ 244
+ ],
+ "score": 1.0,
+ "content": "five framing lines: raise NotImplementedError , pass , assert False , return False , and",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 104,
+ 241,
+ 506,
+ 255
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 241,
+ 506,
+ 255
+ ],
+ "score": 1.0,
+ "content": "print(\"Hello world!\") . We first check that prepending these irrelevant preceding functions",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 506,
+ 267
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 506,
+ 267
+ ],
+ "score": 1.0,
+ "content": "decreases functional accuracy.3 Next, to test if models relied on irrelevant information in the prompt,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 105,
+ 263,
+ 499,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 263,
+ 499,
+ 277
+ ],
+ "score": 1.0,
+ "content": "we measure how much more frequently the framing line appears verbatim in the generated output.",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 357
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "We report the results of our framing experiments in Table 1. We find that adding irrelevant preceding",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "functions consistently lowers functional accuracy, by between 22.3 and 30.5 points for Codex, across",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 302,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "the different framing lines we tested. Moreover, both models frequently generate the framing line:",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 127,
+ 324
+ ],
+ "score": 0.86,
+ "content": "81 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 127,
+ 313,
+ 240,
+ 326
+ ],
+ "score": 1.0,
+ "content": "of the time for Codex and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 313,
+ 268,
+ 324
+ ],
+ "score": 0.88,
+ "content": "7 0 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 313,
+ 439,
+ 326
+ ],
+ "score": 1.0,
+ "content": "of time for CodeGen, compared to only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 439,
+ 313,
+ 462,
+ 324
+ ],
+ "score": 0.87,
+ "content": "4 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 313,
+ 482,
+ 326
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 313,
+ 505,
+ 324
+ ],
+ "score": 0.85,
+ "content": "0 . 0 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "over untransformed prompts respectively. These results suggest that code generation models can",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "erroneously rely on irrelevant information in the prompt in predictable ways, even in the extreme",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 346,
+ 497,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 346,
+ 497,
+ 358
+ ],
+ "score": 1.0,
+ "content": "case when doing so contradicts the type specification in the function signature (return False ).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 370,
+ 238,
+ 382
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 239,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 239,
+ 386
+ ],
+ "score": 1.0,
+ "content": "3.3.2 Inspiration: Anchoring",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 389,
+ 505,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "We next draw inspiration from anchoring: humans’ tendency to insufficiently adjust their estimates",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 400,
+ 505,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 400,
+ 368,
+ 413
+ ],
+ "score": 1.0,
+ "content": "away from initial values. For example, Tversky and Kahneman",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 368,
+ 400,
+ 396,
+ 412
+ ],
+ "score": 0.57,
+ "content": "\\mathbb { \\underline { { \\lVert \\mathbf { 9 7 4 } } \\rVert } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 400,
+ 505,
+ 413
+ ],
+ "score": 1.0,
+ "content": "find that subjects’ median",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 505,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 377,
+ 423
+ ],
+ "score": 1.0,
+ "content": "estimate for the fraction of African countries in the UN shifts from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 378,
+ 412,
+ 397,
+ 422
+ ],
+ "score": 0.83,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 411,
+ 408,
+ 423
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 412,
+ 428,
+ 422
+ ],
+ "score": 0.87,
+ "content": "45 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 411,
+ 505,
+ 423
+ ],
+ "score": 1.0,
+ "content": ", based on whether",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 345,
+ 436
+ ],
+ "score": 1.0,
+ "content": "they were first asked if the fraction was greater or less than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 423,
+ 365,
+ 433
+ ],
+ "score": 0.88,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 420,
+ 383,
+ 436
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 423,
+ 403,
+ 433
+ ],
+ "score": 0.87,
+ "content": "65 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 420,
+ 506,
+ 436
+ ],
+ "score": 1.0,
+ "content": ", respectively. Anchoring",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 497,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 497,
+ 446
+ ],
+ "score": 1.0,
+ "content": "captures how humans adjust to partial information, versus irrelevant information (framing effect).",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 450,
+ 505,
+ 516
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "Using anchoring as inspiration, we hypothesize that code generation models may adjust their output",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "towards related solutions, when these solutions are included in the prompt. To elicit this failure, we",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 471,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 471,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "prepend anchor functions to prompts: functions that are similar to a valid solution for a HumanEval",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 506,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 506,
+ 495
+ ],
+ "score": 1.0,
+ "content": "prompt, but contain some error. We first check that prepending these anchor functions decreases",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 243,
+ 506
+ ],
+ "score": 1.0,
+ "content": "functional accuracy, as in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 493,
+ 269,
+ 506
+ ],
+ "score": 0.48,
+ "content": "3 . 3 . 1 .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 270,
+ 493,
+ 505,
+ 506
+ ],
+ "score": 1.0,
+ "content": "Next, to test if models adjust their output towards related",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 462,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 462,
+ 516
+ ],
+ "score": 1.0,
+ "content": "solutions, we check that the generated solution contains elements of the anchor function.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 520,
+ 505,
+ 565
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "score": 1.0,
+ "content": "We aim to construct anchor functions that are similar to functions in HumanEval prompts and that",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "compile, but are incorrect. To do so, we take a prefix of the canonical solution, then add additional",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 542,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 331,
+ 555
+ ],
+ "score": 1.0,
+ "content": "anchor lines that produce an incorrect output. See Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 331,
+ 542,
+ 340,
+ 555
+ ],
+ "score": 0.56,
+ "content": "\\triangledown",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 543,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "for an example. We describe two types of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 507,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 507,
+ 567
+ ],
+ "score": 1.0,
+ "content": "anchor lines, and how we test their influence on the generated solutions, in the following paragraphs.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 570,
+ 505,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Print-var anchor lines. We first study print-var anchor lines, which iterate over all variables in the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "function signature and print their values. For a function with inputs var1 and var2 , the associated",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 212,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 212,
+ 603
+ ],
+ "score": 1.0,
+ "content": "print-var anchor lines are:",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 608,
+ 232,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 233,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 233,
+ 620
+ ],
+ "score": 1.0,
+ "content": "for var in [var1, var2]:",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 126,
+ 620,
+ 181,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 620,
+ 181,
+ 631
+ ],
+ "score": 1.0,
+ "content": "print(var)",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 636,
+ 504,
+ 659
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "To study the influence of the print-var anchor lines on the solution, we measure how often (i) just the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 497,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 497,
+ 660
+ ],
+ "score": 1.0,
+ "content": "first line (for loop), and (ii) just the second line (print statement) appear in the generated solution.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 663,
+ 505,
+ 696
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "Add-var anchor lines. We also study add-var anchor lines, which return the sum of all variables in",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 686
+ ],
+ "score": 1.0,
+ "content": "the function signature (converted to strings). For a function with inputs var1 and var2 , the add-var",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 684,
+ 176,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 176,
+ 698
+ ],
+ "score": 1.0,
+ "content": "anchor lines are:",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45
+ }
+ ],
+ "page_idx": 4,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 105,
+ 700,
+ 504,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 698,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 698,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "3 Following Chen et al. [2021], we measure performance on HumanEval with functional accuracy: the fraction",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 711,
+ 502,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 711,
+ 502,
+ 723
+ ],
+ "score": 1.0,
+ "content": "of programs that pass all of the test cases provided at the url: https://github.com/openai/human-eval.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 753
+ ],
+ "score": 1.0,
+ "content": "5",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 126
+ ],
+ "score": 0.937,
+ "type": "image",
+ "image_path": "3ca12ceac2dfdc76f5b5cb25925443fec2d6975fface7597f9f0a393592f8cfd.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 109,
+ 69,
+ 502,
+ 88.0
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 109,
+ 88.0,
+ 502,
+ 107.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 109,
+ 107.0,
+ 502,
+ 126.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 197
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 130,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "Figure 3: Illustration of our anchoring experiment using a real example (expanded in Figure 7). We",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 155
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 142,
+ 506,
+ 155
+ ],
+ "score": 1.0,
+ "content": "construct the anchor function (left) by taking the function signature from the HumanEval prompt",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 153,
+ 506,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 153,
+ 180,
+ 165
+ ],
+ "score": 1.0,
+ "content": "(blue), appending",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 180,
+ 155,
+ 188,
+ 163
+ ],
+ "score": 0.62,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 188,
+ 153,
+ 506,
+ 165
+ ],
+ "score": 1.0,
+ "content": "lines of the canonical solution (green), then adding anchoring lines (red). We",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 506,
+ 177
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 506,
+ 177
+ ],
+ "score": 1.0,
+ "content": "construct the full prompt (center) by combining the anchor function, the original HumanEval prompt,",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 106,
+ 176,
+ 505,
+ 187
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 176,
+ 154,
+ 187
+ ],
+ "score": 1.0,
+ "content": "and the first",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 154,
+ 177,
+ 162,
+ 185
+ ],
+ "score": 0.75,
+ "content": "n",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 162,
+ 176,
+ 505,
+ 187
+ ],
+ "score": 1.0,
+ "content": "lines of the canonical solution. The solution Codex generates (right) combines elements",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 185,
+ 501,
+ 199
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 185,
+ 501,
+ 199
+ ],
+ "score": 1.0,
+ "content": "of a canonical solution (checks condition and adds to ret.), with the anchor function (for var loop).",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 5.5
+ }
+ ],
+ "index": 3.25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 219,
+ 505,
+ 275
+ ],
+ "lines": [],
+ "index": 11,
+ "bbox_fs": [
+ 104,
+ 221,
+ 506,
+ 277
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 280,
+ 505,
+ 357
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 294
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 279,
+ 505,
+ 294
+ ],
+ "score": 1.0,
+ "content": "We report the results of our framing experiments in Table 1. We find that adding irrelevant preceding",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 106,
+ 291,
+ 506,
+ 304
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 291,
+ 506,
+ 304
+ ],
+ "score": 1.0,
+ "content": "functions consistently lowers functional accuracy, by between 22.3 and 30.5 points for Codex, across",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 106,
+ 302,
+ 506,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 302,
+ 506,
+ 315
+ ],
+ "score": 1.0,
+ "content": "the different framing lines we tested. Moreover, both models frequently generate the framing line:",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 106,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 313,
+ 127,
+ 324
+ ],
+ "score": 0.86,
+ "content": "81 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 127,
+ 313,
+ 240,
+ 326
+ ],
+ "score": 1.0,
+ "content": "of the time for Codex and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 313,
+ 268,
+ 324
+ ],
+ "score": 0.88,
+ "content": "7 0 . 7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 269,
+ 313,
+ 439,
+ 326
+ ],
+ "score": 1.0,
+ "content": "of time for CodeGen, compared to only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 439,
+ 313,
+ 462,
+ 324
+ ],
+ "score": 0.87,
+ "content": "4 . 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 462,
+ 313,
+ 482,
+ 326
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 482,
+ 313,
+ 505,
+ 324
+ ],
+ "score": 0.85,
+ "content": "0 . 0 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 337
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 324,
+ 505,
+ 337
+ ],
+ "score": 1.0,
+ "content": "over untransformed prompts respectively. These results suggest that code generation models can",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 336,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "erroneously rely on irrelevant information in the prompt in predictable ways, even in the extreme",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 346,
+ 497,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 346,
+ 497,
+ 358
+ ],
+ "score": 1.0,
+ "content": "case when doing so contradicts the type specification in the function signature (return False ).",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 279,
+ 506,
+ 358
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 370,
+ 238,
+ 382
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 239,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 367,
+ 239,
+ 386
+ ],
+ "score": 1.0,
+ "content": "3.3.2 Inspiration: Anchoring",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 389,
+ 505,
+ 445
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 505,
+ 401
+ ],
+ "score": 1.0,
+ "content": "We next draw inspiration from anchoring: humans’ tendency to insufficiently adjust their estimates",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 400,
+ 505,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 400,
+ 368,
+ 413
+ ],
+ "score": 1.0,
+ "content": "away from initial values. For example, Tversky and Kahneman",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 368,
+ 400,
+ 396,
+ 412
+ ],
+ "score": 0.57,
+ "content": "\\mathbb { \\underline { { \\lVert \\mathbf { 9 7 4 } } \\rVert } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 396,
+ 400,
+ 505,
+ 413
+ ],
+ "score": 1.0,
+ "content": "find that subjects’ median",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 411,
+ 505,
+ 423
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 411,
+ 377,
+ 423
+ ],
+ "score": 1.0,
+ "content": "estimate for the fraction of African countries in the UN shifts from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 378,
+ 412,
+ 397,
+ 422
+ ],
+ "score": 0.83,
+ "content": "2 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 398,
+ 411,
+ 408,
+ 423
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 408,
+ 412,
+ 428,
+ 422
+ ],
+ "score": 0.87,
+ "content": "45 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 429,
+ 411,
+ 505,
+ 423
+ ],
+ "score": 1.0,
+ "content": ", based on whether",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 420,
+ 506,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 420,
+ 345,
+ 436
+ ],
+ "score": 1.0,
+ "content": "they were first asked if the fraction was greater or less than",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 345,
+ 423,
+ 365,
+ 433
+ ],
+ "score": 0.88,
+ "content": "10 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 365,
+ 420,
+ 383,
+ 436
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 383,
+ 423,
+ 403,
+ 433
+ ],
+ "score": 0.87,
+ "content": "65 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 403,
+ 420,
+ 506,
+ 436
+ ],
+ "score": 1.0,
+ "content": ", respectively. Anchoring",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 433,
+ 497,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 433,
+ 497,
+ 446
+ ],
+ "score": 1.0,
+ "content": "captures how humans adjust to partial information, versus irrelevant information (framing effect).",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 389,
+ 506,
+ 446
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 450,
+ 505,
+ 516
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "Using anchoring as inspiration, we hypothesize that code generation models may adjust their output",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "towards related solutions, when these solutions are included in the prompt. To elicit this failure, we",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 104,
+ 471,
+ 506,
+ 484
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 471,
+ 506,
+ 484
+ ],
+ "score": 1.0,
+ "content": "prepend anchor functions to prompts: functions that are similar to a valid solution for a HumanEval",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 506,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 506,
+ 495
+ ],
+ "score": 1.0,
+ "content": "prompt, but contain some error. We first check that prepending these anchor functions decreases",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 493,
+ 505,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 493,
+ 243,
+ 506
+ ],
+ "score": 1.0,
+ "content": "functional accuracy, as in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 244,
+ 493,
+ 269,
+ 506
+ ],
+ "score": 0.48,
+ "content": "3 . 3 . 1 .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 270,
+ 493,
+ 505,
+ 506
+ ],
+ "score": 1.0,
+ "content": "Next, to test if models adjust their output towards related",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 505,
+ 462,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 505,
+ 462,
+ 516
+ ],
+ "score": 1.0,
+ "content": "solutions, we check that the generated solution contains elements of the anchor function.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 29.5,
+ "bbox_fs": [
+ 104,
+ 449,
+ 506,
+ 516
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 520,
+ 505,
+ 565
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 521,
+ 504,
+ 532
+ ],
+ "score": 1.0,
+ "content": "We aim to construct anchor functions that are similar to functions in HumanEval prompts and that",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 543
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 543
+ ],
+ "score": 1.0,
+ "content": "compile, but are incorrect. To do so, we take a prefix of the canonical solution, then add additional",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 542,
+ 505,
+ 555
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 543,
+ 331,
+ 555
+ ],
+ "score": 1.0,
+ "content": "anchor lines that produce an incorrect output. See Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 331,
+ 542,
+ 340,
+ 555
+ ],
+ "score": 0.56,
+ "content": "\\triangledown",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 341,
+ 543,
+ 505,
+ 555
+ ],
+ "score": 1.0,
+ "content": "for an example. We describe two types of",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 552,
+ 507,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 552,
+ 507,
+ 567
+ ],
+ "score": 1.0,
+ "content": "anchor lines, and how we test their influence on the generated solutions, in the following paragraphs.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 34.5,
+ "bbox_fs": [
+ 105,
+ 521,
+ 507,
+ 567
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 570,
+ 505,
+ 603
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 570,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "Print-var anchor lines. We first study print-var anchor lines, which iterate over all variables in the",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 581,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "function signature and print their values. For a function with inputs var1 and var2 , the associated",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 592,
+ 212,
+ 603
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 592,
+ 212,
+ 603
+ ],
+ "score": 1.0,
+ "content": "print-var anchor lines are:",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 570,
+ 505,
+ 603
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 608,
+ 232,
+ 631
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 233,
+ 620
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 608,
+ 233,
+ 620
+ ],
+ "score": 1.0,
+ "content": "for var in [var1, var2]:",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 126,
+ 620,
+ 181,
+ 631
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 126,
+ 620,
+ 181,
+ 631
+ ],
+ "score": 1.0,
+ "content": "print(var)",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 106,
+ 608,
+ 233,
+ 631
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 636,
+ 504,
+ 659
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 636,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "To study the influence of the print-var anchor lines on the solution, we measure how often (i) just the",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 646,
+ 497,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 646,
+ 497,
+ 660
+ ],
+ "score": 1.0,
+ "content": "first line (for loop), and (ii) just the second line (print statement) appear in the generated solution.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42.5,
+ "bbox_fs": [
+ 105,
+ 636,
+ 505,
+ 660
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 663,
+ 505,
+ 696
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 663,
+ 505,
+ 675
+ ],
+ "score": 1.0,
+ "content": "Add-var anchor lines. We also study add-var anchor lines, which return the sum of all variables in",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 686
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 675,
+ 505,
+ 686
+ ],
+ "score": 1.0,
+ "content": "the function signature (converted to strings). For a function with inputs var1 and var2 , the add-var",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 684,
+ 176,
+ 698
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 176,
+ 698
+ ],
+ "score": 1.0,
+ "content": "anchor lines are:",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ }
+ ],
+ "index": 45,
+ "bbox_fs": [
+ 105,
+ 663,
+ 505,
+ 698
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "5fd2c0df7eb36902206a8d073de7ef1e8a2e8431015b1607d9f4d14ef1f86fae.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 130.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 130.66666666666666,
+ 499,
+ 189.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 189.33333333333331,
+ 499,
+ 247.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 256,
+ 506,
+ 344
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Results of the print-var anchoring experiment. Left. We measure the functional accuracy of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 268,
+ 507,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 507,
+ 280
+ ],
+ "score": 1.0,
+ "content": "Codex (top) and CodeGen (bottom) with no anchor function prepended (baseline acc) and with a print-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 279,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 279,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "var anchor function prepended (anchor acc), and find that prepending the anchor function consistently",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": "lowers accuracy. Right. We measure the influence of the anchor function on the generated solution",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 506,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 506,
+ 313
+ ],
+ "score": 1.0,
+ "content": "by plotting the fraction of generated solutions that contain “for var in ” from the print-var anchor",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "score": 1.0,
+ "content": "prompt (for var loop), the fraction of generated solutions that include “print(var) ” (prints var),",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "and the fraction of generated solutions that output the anchor function verbatim without additional",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 497,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 497,
+ 346
+ ],
+ "score": 1.0,
+ "content": "content (exact copy), as a function of the number of canonical solution lines added to the prompt.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 365,
+ 248,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 249,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 189,
+ 377
+ ],
+ "score": 1.0,
+ "content": "tmp = str(var1)",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 366,
+ 198,
+ 375
+ ],
+ "score": 0.33,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 364,
+ 249,
+ 377
+ ],
+ "score": 1.0,
+ "content": "str(var2)",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 162,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 162,
+ 389
+ ],
+ "score": 1.0,
+ "content": "return tmp",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 392,
+ 504,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "To study the influence of the add-var anchor lines on the solution, we measure how often return",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 260,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 260,
+ 415
+ ],
+ "score": 1.0,
+ "content": "tmp appears in the generated solution.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 419,
+ 505,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 418,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 418,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "Print-var results. In Figure 4, we show that prepending print-var anchor functions consistently lowers",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "Codex and CodeGens’ functional accuracies across different number of prompted canonical solution",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "lines. We vary the number of canonical solution lines to study prompts of different difficulties; as the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 495,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 495,
+ 464
+ ],
+ "score": 1.0,
+ "content": "number of solution lines increases, the number remaining lines models must produce decreases.4",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 546
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "We additionally find that elements of anchor function often appear in both models’ outputs, suggesting",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 479,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 447,
+ 492
+ ],
+ "score": 1.0,
+ "content": "that code generation models adjust their solutions towards related solutions. In Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 479,
+ 457,
+ 492
+ ],
+ "score": 0.51,
+ "content": "^ { 4 , }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 458,
+ 480,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "we see that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 491,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 491,
+ 222,
+ 502
+ ],
+ "score": 1.0,
+ "content": "Codex generates for var in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 491,
+ 264,
+ 501
+ ],
+ "score": 0.88,
+ "content": "3 2 \\% { - } 6 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 491,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "of solutions when at least one line of the canonical solution is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 266,
+ 514
+ ],
+ "score": 1.0,
+ "content": "included, and generates print(var) in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 267,
+ 502,
+ 309,
+ 512
+ ],
+ "score": 0.89,
+ "content": "2 6 \\% - 4 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 501,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "of solutions. CodeGen’s behavior is qualitatively",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "similar. Both models sometimes even incorporate the anchor lines into correct solutions; on Codex,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 302,
+ 536
+ ],
+ "score": 1.0,
+ "content": "the for var loop is used in a correct solution for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 523,
+ 340,
+ 534
+ ],
+ "score": 0.87,
+ "content": "3 \\% - 1 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 523,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "of all outputs, while print(var) is used",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 281,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 203,
+ 547
+ ],
+ "score": 1.0,
+ "content": "in a correct solution for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 534,
+ 235,
+ 545
+ ],
+ "score": 0.9,
+ "content": "1 \\% - 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 534,
+ 281,
+ 547
+ ],
+ "score": 1.0,
+ "content": "of outputs.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 550,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "score": 1.0,
+ "content": "Control experiments. One concern might be that models just outputs the anchor function verbatim,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 157,
+ 573
+ ],
+ "score": 1.0,
+ "content": "as in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 561,
+ 182,
+ 574
+ ],
+ "score": 0.87,
+ "content": "\\underline { { \\left. 3 . 3 . 1 \\right. } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 562,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "but we find that this does not explain the full results—both models include anchor",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "lines in many solutions that do not copy the anchor function verbatim. We also find that changing the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 474,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 474,
+ 596
+ ],
+ "score": 1.0,
+ "content": "name of the anchor function leads to only negligible changes; see Appendix A.1 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 504,
+ 623
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "Add-var results. We next consider results for add-var anchor lines. Full results for the add-var",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 610,
+ 502,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 277,
+ 623
+ ],
+ "score": 1.0,
+ "content": "anchor prompts are presented in Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 610,
+ 296,
+ 623
+ ],
+ "score": 0.48,
+ "content": "\\mathbf { A . l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 610,
+ 502,
+ 623
+ ],
+ "score": 1.0,
+ "content": "and are qualitatively similar to the print-var results.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 627,
+ 505,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 507,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 507,
+ 640
+ ],
+ "score": 1.0,
+ "content": "One again, we find that prepending the anchor function consistently lowers functional accuracy.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 638,
+ 504,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 504,
+ 649
+ ],
+ "score": 1.0,
+ "content": "Moreover, the outputted solutions often include an anchor line. For example, Codex and CodeGen",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 648,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 648,
+ 206,
+ 663
+ ],
+ "score": 1.0,
+ "content": "generate return tmp in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 649,
+ 248,
+ 660
+ ],
+ "score": 0.89,
+ "content": "2 6 \\% - 4 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 648,
+ 266,
+ 663
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 649,
+ 308,
+ 659
+ ],
+ "score": 0.9,
+ "content": "1 3 \\% - 7 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 648,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": "of solutions respectively, depending on how many",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "canonical solution lines we prompt with. These results are not caused by models outputting the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 327,
+ 683
+ ],
+ "score": 1.0,
+ "content": "anchoring function verbatim: this only occurs between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 671,
+ 343,
+ 681
+ ],
+ "score": 0.85,
+ "content": "7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 671,
+ 361,
+ 683
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 671,
+ 381,
+ 681
+ ],
+ "score": 0.87,
+ "content": "12 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 671,
+ 489,
+ 683
+ ],
+ "score": 1.0,
+ "content": "of the time for Codex, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 671,
+ 505,
+ 681
+ ],
+ "score": 0.84,
+ "content": "4 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 680,
+ 506,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 680,
+ 123,
+ 695
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 123,
+ 682,
+ 142,
+ 692
+ ],
+ "score": 0.87,
+ "content": "12 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 680,
+ 506,
+ 695
+ ],
+ "score": 1.0,
+ "content": "for CodeGen. Overall, our findings suggest that code generation models can err by adjusting",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 692,
+ 438,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 692,
+ 438,
+ 705
+ ],
+ "score": 1.0,
+ "content": "its output towards related solutions, when the solutions are included in the prompt.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "page_idx": 5,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 114,
+ 711,
+ 491,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 708,
+ 493,
+ 725
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 708,
+ 493,
+ 725
+ ],
+ "score": 1.0,
+ "content": "4 We filter out programs whose entire canonical solution would appear in the prompt; see Appendix A.1.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 309,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "6",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 333,
+ 74,
+ 490,
+ 81
+ ],
+ "lines": []
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "image",
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 248
+ ],
+ "score": 0.972,
+ "type": "image",
+ "image_path": "5fd2c0df7eb36902206a8d073de7ef1e8a2e8431015b1607d9f4d14ef1f86fae.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 110,
+ 72,
+ 499,
+ 130.66666666666666
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 110,
+ 130.66666666666666,
+ 499,
+ 189.33333333333331
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 110,
+ 189.33333333333331,
+ 499,
+ 247.99999999999997
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 256,
+ 506,
+ 344
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 506,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 257,
+ 506,
+ 268
+ ],
+ "score": 1.0,
+ "content": "Figure 4: Results of the print-var anchoring experiment. Left. We measure the functional accuracy of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 106,
+ 268,
+ 507,
+ 280
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 268,
+ 507,
+ 280
+ ],
+ "score": 1.0,
+ "content": "Codex (top) and CodeGen (bottom) with no anchor function prepended (baseline acc) and with a print-",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 106,
+ 279,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 279,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "var anchor function prepended (anchor acc), and find that prepending the anchor function consistently",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 289,
+ 505,
+ 302
+ ],
+ "score": 1.0,
+ "content": "lowers accuracy. Right. We measure the influence of the anchor function on the generated solution",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 506,
+ 313
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 506,
+ 313
+ ],
+ "score": 1.0,
+ "content": "by plotting the fraction of generated solutions that contain “for var in ” from the print-var anchor",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 506,
+ 324
+ ],
+ "score": 1.0,
+ "content": "prompt (for var loop), the fraction of generated solutions that include “print(var) ” (prints var),",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 506,
+ 334
+ ],
+ "score": 1.0,
+ "content": "and the fraction of generated solutions that output the anchor function verbatim without additional",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 497,
+ 346
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 497,
+ 346
+ ],
+ "score": 1.0,
+ "content": "content (exact copy), as a function of the number of canonical solution lines added to the prompt.",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ }
+ ],
+ "index": 6.5
+ }
+ ],
+ "index": 3.75
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 365,
+ 248,
+ 387
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 249,
+ 377
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 364,
+ 189,
+ 377
+ ],
+ "score": 1.0,
+ "content": "tmp = str(var1)",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 189,
+ 366,
+ 198,
+ 375
+ ],
+ "score": 0.33,
+ "content": "^ +",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 198,
+ 364,
+ 249,
+ 377
+ ],
+ "score": 1.0,
+ "content": "str(var2)",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 105,
+ 375,
+ 162,
+ 389
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 375,
+ 162,
+ 389
+ ],
+ "score": 1.0,
+ "content": "return tmp",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 105,
+ 364,
+ 249,
+ 389
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 392,
+ 504,
+ 414
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 505,
+ 405
+ ],
+ "score": 1.0,
+ "content": "To study the influence of the add-var anchor lines on the solution, we measure how often return",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 404,
+ 260,
+ 415
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 404,
+ 260,
+ 415
+ ],
+ "score": 1.0,
+ "content": "tmp appears in the generated solution.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 106,
+ 392,
+ 505,
+ 415
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 419,
+ 505,
+ 463
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 418,
+ 506,
+ 433
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 418,
+ 506,
+ 433
+ ],
+ "score": 1.0,
+ "content": "Print-var results. In Figure 4, we show that prepending print-var anchor functions consistently lowers",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 430,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": "Codex and CodeGens’ functional accuracies across different number of prompted canonical solution",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 441,
+ 505,
+ 454
+ ],
+ "score": 1.0,
+ "content": "lines. We vary the number of canonical solution lines to study prompts of different difficulties; as the",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 495,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 495,
+ 464
+ ],
+ "score": 1.0,
+ "content": "number of solution lines increases, the number remaining lines models must produce decreases.4",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 104,
+ 418,
+ 506,
+ 464
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 469,
+ 505,
+ 546
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 482
+ ],
+ "score": 1.0,
+ "content": "We additionally find that elements of anchor function often appear in both models’ outputs, suggesting",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 106,
+ 479,
+ 505,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 480,
+ 447,
+ 492
+ ],
+ "score": 1.0,
+ "content": "that code generation models adjust their solutions towards related solutions. In Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 447,
+ 479,
+ 457,
+ 492
+ ],
+ "score": 0.51,
+ "content": "^ { 4 , }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 458,
+ 480,
+ 505,
+ 492
+ ],
+ "score": 1.0,
+ "content": "we see that",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 106,
+ 491,
+ 506,
+ 502
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 491,
+ 222,
+ 502
+ ],
+ "score": 1.0,
+ "content": "Codex generates for var in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 222,
+ 491,
+ 264,
+ 501
+ ],
+ "score": 0.88,
+ "content": "3 2 \\% { - } 6 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 491,
+ 506,
+ 502
+ ],
+ "score": 1.0,
+ "content": "of solutions when at least one line of the canonical solution is",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 501,
+ 505,
+ 514
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 501,
+ 266,
+ 514
+ ],
+ "score": 1.0,
+ "content": "included, and generates print(var) in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 267,
+ 502,
+ 309,
+ 512
+ ],
+ "score": 0.89,
+ "content": "2 6 \\% - 4 4 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 309,
+ 501,
+ 505,
+ 514
+ ],
+ "score": 1.0,
+ "content": "of solutions. CodeGen’s behavior is qualitatively",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 511,
+ 506,
+ 525
+ ],
+ "score": 1.0,
+ "content": "similar. Both models sometimes even incorporate the anchor lines into correct solutions; on Codex,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 523,
+ 506,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 523,
+ 302,
+ 536
+ ],
+ "score": 1.0,
+ "content": "the for var loop is used in a correct solution for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 523,
+ 340,
+ 534
+ ],
+ "score": 0.87,
+ "content": "3 \\% - 1 1 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 523,
+ 506,
+ 536
+ ],
+ "score": 1.0,
+ "content": "of all outputs, while print(var) is used",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 534,
+ 281,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 534,
+ 203,
+ 547
+ ],
+ "score": 1.0,
+ "content": "in a correct solution for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 203,
+ 534,
+ 235,
+ 545
+ ],
+ "score": 0.9,
+ "content": "1 \\% - 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 534,
+ 281,
+ 547
+ ],
+ "score": 1.0,
+ "content": "of outputs.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 22,
+ "bbox_fs": [
+ 105,
+ 468,
+ 506,
+ 547
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 550,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 550,
+ 506,
+ 563
+ ],
+ "score": 1.0,
+ "content": "Control experiments. One concern might be that models just outputs the anchor function verbatim,",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 106,
+ 561,
+ 506,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 157,
+ 573
+ ],
+ "score": 1.0,
+ "content": "as in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 158,
+ 561,
+ 182,
+ 574
+ ],
+ "score": 0.87,
+ "content": "\\underline { { \\left. 3 . 3 . 1 \\right. } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 182,
+ 562,
+ 506,
+ 573
+ ],
+ "score": 1.0,
+ "content": "but we find that this does not explain the full results—both models include anchor",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "lines in many solutions that do not copy the anchor function verbatim. We also find that changing the",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 474,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 474,
+ 596
+ ],
+ "score": 1.0,
+ "content": "name of the anchor function leads to only negligible changes; see Appendix A.1 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 27.5,
+ "bbox_fs": [
+ 105,
+ 550,
+ 506,
+ 596
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 504,
+ 623
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 611
+ ],
+ "score": 1.0,
+ "content": "Add-var results. We next consider results for add-var anchor lines. Full results for the add-var",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 610,
+ 502,
+ 623
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 610,
+ 277,
+ 623
+ ],
+ "score": 1.0,
+ "content": "anchor prompts are presented in Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 277,
+ 610,
+ 296,
+ 623
+ ],
+ "score": 0.48,
+ "content": "\\mathbf { A . l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 297,
+ 610,
+ 502,
+ 623
+ ],
+ "score": 1.0,
+ "content": "and are qualitatively similar to the print-var results.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 600,
+ 505,
+ 623
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 627,
+ 505,
+ 704
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 507,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 507,
+ 640
+ ],
+ "score": 1.0,
+ "content": "One again, we find that prepending the anchor function consistently lowers functional accuracy.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 106,
+ 638,
+ 504,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 504,
+ 649
+ ],
+ "score": 1.0,
+ "content": "Moreover, the outputted solutions often include an anchor line. For example, Codex and CodeGen",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 104,
+ 648,
+ 506,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 648,
+ 206,
+ 663
+ ],
+ "score": 1.0,
+ "content": "generate return tmp in",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 207,
+ 649,
+ 248,
+ 660
+ ],
+ "score": 0.89,
+ "content": "2 6 \\% - 4 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 249,
+ 648,
+ 266,
+ 663
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 266,
+ 649,
+ 308,
+ 659
+ ],
+ "score": 0.9,
+ "content": "1 3 \\% - 7 9 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 308,
+ 648,
+ 506,
+ 663
+ ],
+ "score": 1.0,
+ "content": "of solutions respectively, depending on how many",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 660,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "canonical solution lines we prompt with. These results are not caused by models outputting the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 671,
+ 505,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 671,
+ 327,
+ 683
+ ],
+ "score": 1.0,
+ "content": "anchoring function verbatim: this only occurs between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 328,
+ 671,
+ 343,
+ 681
+ ],
+ "score": 0.85,
+ "content": "7 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 343,
+ 671,
+ 361,
+ 683
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 361,
+ 671,
+ 381,
+ 681
+ ],
+ "score": 0.87,
+ "content": "12 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 381,
+ 671,
+ 489,
+ 683
+ ],
+ "score": 1.0,
+ "content": "of the time for Codex, and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 490,
+ 671,
+ 505,
+ 681
+ ],
+ "score": 0.84,
+ "content": "4 \\%",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 680,
+ 506,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 680,
+ 123,
+ 695
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 123,
+ 682,
+ 142,
+ 692
+ ],
+ "score": 0.87,
+ "content": "12 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 142,
+ 680,
+ 506,
+ 695
+ ],
+ "score": 1.0,
+ "content": "for CodeGen. Overall, our findings suggest that code generation models can err by adjusting",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 692,
+ 438,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 692,
+ 438,
+ 705
+ ],
+ "score": 1.0,
+ "content": "its output towards related solutions, when the solutions are included in the prompt.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 104,
+ 626,
+ 507,
+ 705
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 83,
+ 286,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 82,
+ 285,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 82,
+ 285,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Write a function that squares",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 113,
+ 95,
+ 239,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 95,
+ 239,
+ 106
+ ],
+ "score": 1.0,
+ "content": "the sum of its inputs",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 113,
+ 106,
+ 133,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 106,
+ 133,
+ 115
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 310,
+ 84,
+ 489,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 82,
+ 488,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 82,
+ 488,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Write a function that sums its",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 310,
+ 94,
+ 477,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 94,
+ 477,
+ 107
+ ],
+ "score": 1.0,
+ "content": "inputs called product_plus_2",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 310,
+ 106,
+ 330,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 106,
+ 330,
+ 115
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 120,
+ 273,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 118,
+ 239,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 118,
+ 239,
+ 132
+ ],
+ "score": 1.0,
+ "content": "def square_sum(x, y):",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 129,
+ 272,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 129,
+ 183,
+ 142
+ ],
+ "score": 1.0,
+ "content": "return",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 131,
+ 272,
+ 142
+ ],
+ "score": 0.55,
+ "content": "\\texttt { x } \\star \\texttt { 2 } + \\texttt { y } \\star \\texttt { 2 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 148,
+ 506,
+ 214
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Left. Availability heuristic example where Codex mixes up the order of operations. The",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "correct function signature (blue), square_sum matches the prompt. However, the incorrect function",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "call (red) instead squares its inputs before summing them. The prompt is above the horizontal line,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 193
+ ],
+ "score": 1.0,
+ "content": "while the generated code is below. Right. Attribute substitution example where Codex relies on the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 191,
+ 506,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 191,
+ 506,
+ 204
+ ],
+ "score": 1.0,
+ "content": "function name to generate output. Codex correctly generates the desired function name (blue), but",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 203,
+ 474,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 203,
+ 474,
+ 216
+ ],
+ "score": 1.0,
+ "content": "errs by using the function name instead of the prompt to generate the return statement (red).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 234,
+ 281,
+ 246
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 281,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 281,
+ 248
+ ],
+ "score": 1.0,
+ "content": "3.3.3 Inspiration: Availability heuristic",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 253,
+ 505,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "We next draw inspiration from the availability heuristic: the tendency of humans to evaluate how",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 276
+ ],
+ "score": 1.0,
+ "content": "frequently an example occurs based on how easy it is to recall. For example, Tversky and Kahneman",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 107,
+ 275,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 275,
+ 135,
+ 287
+ ],
+ "score": 0.56,
+ "content": "\\mathbb { \\underline { { \\lVert \\nabla ^ { 9 } 7 3 \\rVert } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 275,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "find that humans tend to incorrectly report that there are more first words that start with “r”",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 446,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 446,
+ 299
+ ],
+ "score": 1.0,
+ "content": "and “k” than have third letter “r” and “k”, because the former quickly come to mind.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 302,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Using the availability heuristic as motivation, we hypothesize that code generation models may err by",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "outputting solutions to related prompts that appear more frequently in the training set. To elicit this",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "failure, we start with prompts that apply a unary operation before a binary operation (unary-first), then",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 349
+ ],
+ "score": 1.0,
+ "content": "flip the order (binary-first). Programmers tend to apply unary operations first (e.g. when computing",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 346,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 346,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Euclidean distances or variances), so we conjecture that they appear more frequently on GitHub. We",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "first check that flipping the order of operations decreases accuracy. Next, to test if code generation",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "models instead outputs related prompts that occur more frequently in the training set, we measure",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 379,
+ 392,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 392,
+ 391
+ ],
+ "score": 1.0,
+ "content": "whether code generation models instead output the unary-first solution.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 395,
+ 505,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "We consider all 12 combinations of the binary operations sum, difference, and product, with unary",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "operations square, cube, quadruple, and square root. Focusing on Codex,5 we find that accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 151,
+ 429
+ ],
+ "score": 1.0,
+ "content": "drops from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 152,
+ 417,
+ 171,
+ 428
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 417,
+ 182,
+ 429
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 417,
+ 201,
+ 428
+ ],
+ "score": 0.87,
+ "content": "17 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "when flipping the order from unary-first to binary-first. Among combinations",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 309,
+ 441
+ ],
+ "score": 1.0,
+ "content": "where flipping the order leads to error, we find that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 428,
+ 329,
+ 439
+ ],
+ "score": 0.88,
+ "content": "7 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 428,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "of the binary-first outputs are the unary-first",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 281,
+ 452
+ ],
+ "score": 1.0,
+ "content": "solution. We exhibit one such error in Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 438,
+ 292,
+ 451
+ ],
+ "score": 0.46,
+ "content": "\\boxed { 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 439,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "when prompted to square the sum of its inputs, Codex",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "generates the correct function name (square_sum ), but reverses the order of operations. Our results",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 497,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 497,
+ 474
+ ],
+ "score": 1.0,
+ "content": "suggest that Codex can err by outputting solutions to related, frequent prompts in the training set.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 477,
+ 505,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 506,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 506,
+ 489
+ ],
+ "score": 1.0,
+ "content": "Control experiments. One worry is that the dip in performance is due the instructional nature of our",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "prompts. We rule this out by evaluating Codex on prompts where the docstring appears beneath the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 511
+ ],
+ "score": 1.0,
+ "content": "function signature and is a definition rather than command, to more closely mimic some functions on",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 509,
+ 487,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 487,
+ 524
+ ],
+ "score": 1.0,
+ "content": "GitHub. We obtain qualitatively similar results on these prompts, see Appendix A.4 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 533,
+ 284,
+ 544
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 285,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 285,
+ 546
+ ],
+ "score": 1.0,
+ "content": "3.3.4 Inspiration: Attribute substitution",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 551,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 506,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 506,
+ 564
+ ],
+ "score": 1.0,
+ "content": "Finally, we draw inspiration from attribute substitution: the human tendency to respond to a com-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 562,
+ 506,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 506,
+ 574
+ ],
+ "score": 1.0,
+ "content": "plicated question using a simpler, related question [Kahneman and Frederick, 2002]. For example,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 104,
+ 573,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 573,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "a professor when asked how likely a candidate is to be tenured, may instead respond with how",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 253,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 253,
+ 595
+ ],
+ "score": 1.0,
+ "content": "impressive they found their job talk.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 505,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "Using attribute substitution as inspiration, we hypothesize that Codex may use simple-but-incorrect",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 611,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 611,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": "heuristics to generate solutions. To elicit this failure, we add requests for conflicting function names",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "to MathEquation prompts. For example, in Figure 5 we prompt Codex to write a program that sums",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "its inputs called product_plus_2 . We first check that adding conflicting function names decreases",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "Codex’s functional accuracy. Next, to test if Codex uses simple-but-incorrect heuristics to generate",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 419,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 419,
+ 667
+ ],
+ "score": 1.0,
+ "content": "solutions, we check whether the generate solution matches the function name.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 671,
+ 503,
+ 694
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "We evaluate Codex using 90 MathEquation prompts where the desired solution and requested",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 682,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 682,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": "function name differ. To construct prompts, we begin with a prompt that Codex originally solves",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49.5
+ }
+ ],
+ "page_idx": 6,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 108,
+ 701,
+ 504,
+ 723
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 506,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 699,
+ 506,
+ 714
+ ],
+ "score": 1.0,
+ "content": "5 We find that CodeGen often produces nonsensical solutions on the style of prompts used in Section 3.3.3,",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 106,
+ 710,
+ 331,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 710,
+ 331,
+ 724
+ ],
+ "score": 1.0,
+ "content": "Section 3.3.4, and Section 5, so we focus primarily on Codex.",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "7",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 114,
+ 72,
+ 131,
+ 78
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 114,
+ 73,
+ 132,
+ 78
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 73,
+ 132,
+ 78
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 311,
+ 72,
+ 329,
+ 78
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 311,
+ 73,
+ 329,
+ 78
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 311,
+ 73,
+ 329,
+ 78
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 311,
+ 119,
+ 457,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 117,
+ 459,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 117,
+ 459,
+ 132
+ ],
+ "score": 1.0,
+ "content": "def product_plus_2(x, y):",
+ "type": "text"
+ }
+ ]
+ },
+ {
+ "bbox": [
+ 333,
+ 130,
+ 430,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 333,
+ 130,
+ 430,
+ 142
+ ],
+ "score": 1.0,
+ "content": "return x * y + 2",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 83,
+ 286,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 82,
+ 285,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 82,
+ 285,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Write a function that squares",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 113,
+ 95,
+ 239,
+ 106
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 95,
+ 239,
+ 106
+ ],
+ "score": 1.0,
+ "content": "the sum of its inputs",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 113,
+ 106,
+ 133,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 113,
+ 106,
+ 133,
+ 115
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 112,
+ 82,
+ 285,
+ 115
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 310,
+ 84,
+ 489,
+ 115
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 309,
+ 82,
+ 488,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 309,
+ 82,
+ 488,
+ 95
+ ],
+ "score": 1.0,
+ "content": "Write a function that sums its",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 310,
+ 94,
+ 477,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 94,
+ 477,
+ 107
+ ],
+ "score": 1.0,
+ "content": "inputs called product_plus_2",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 310,
+ 106,
+ 330,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 310,
+ 106,
+ 330,
+ 115
+ ],
+ "score": 0.0,
+ "content": "",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 309,
+ 82,
+ 488,
+ 115
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 113,
+ 120,
+ 273,
+ 141
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 112,
+ 118,
+ 239,
+ 132
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 112,
+ 118,
+ 239,
+ 132
+ ],
+ "score": 1.0,
+ "content": "def square_sum(x, y):",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 141,
+ 129,
+ 272,
+ 142
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 141,
+ 129,
+ 183,
+ 142
+ ],
+ "score": 1.0,
+ "content": "return",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 183,
+ 131,
+ 272,
+ 142
+ ],
+ "score": 0.55,
+ "content": "\\texttt { x } \\star \\texttt { 2 } + \\texttt { y } \\star \\texttt { 2 }",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 6.5,
+ "bbox_fs": [
+ 112,
+ 118,
+ 272,
+ 142
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 148,
+ 506,
+ 214
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 506,
+ 160
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 147,
+ 506,
+ 160
+ ],
+ "score": 1.0,
+ "content": "Figure 5: Left. Availability heuristic example where Codex mixes up the order of operations. The",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 159,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "correct function signature (blue), square_sum matches the prompt. However, the incorrect function",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 183
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 183
+ ],
+ "score": 1.0,
+ "content": "call (red) instead squares its inputs before summing them. The prompt is above the horizontal line,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 193
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 180,
+ 506,
+ 193
+ ],
+ "score": 1.0,
+ "content": "while the generated code is below. Right. Attribute substitution example where Codex relies on the",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 191,
+ 506,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 191,
+ 506,
+ 204
+ ],
+ "score": 1.0,
+ "content": "function name to generate output. Codex correctly generates the desired function name (blue), but",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 203,
+ 474,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 203,
+ 474,
+ 216
+ ],
+ "score": 1.0,
+ "content": "errs by using the function name instead of the prompt to generate the return statement (red).",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 105,
+ 147,
+ 506,
+ 216
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 234,
+ 281,
+ 246
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 281,
+ 248
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 234,
+ 281,
+ 248
+ ],
+ "score": 1.0,
+ "content": "3.3.3 Inspiration: Availability heuristic",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 14
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 253,
+ 505,
+ 297
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 253,
+ 506,
+ 266
+ ],
+ "score": 1.0,
+ "content": "We next draw inspiration from the availability heuristic: the tendency of humans to evaluate how",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 264,
+ 506,
+ 276
+ ],
+ "score": 1.0,
+ "content": "frequently an example occurs based on how easy it is to recall. For example, Tversky and Kahneman",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 107,
+ 275,
+ 505,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 275,
+ 135,
+ 287
+ ],
+ "score": 0.56,
+ "content": "\\mathbb { \\underline { { \\lVert \\nabla ^ { 9 } 7 3 \\rVert } } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 275,
+ 505,
+ 288
+ ],
+ "score": 1.0,
+ "content": "find that humans tend to incorrectly report that there are more first words that start with “r”",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 286,
+ 446,
+ 299
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 286,
+ 446,
+ 299
+ ],
+ "score": 1.0,
+ "content": "and “k” than have third letter “r” and “k”, because the former quickly come to mind.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 105,
+ 253,
+ 506,
+ 299
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 302,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 505,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 302,
+ 505,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Using the availability heuristic as motivation, we hypothesize that code generation models may err by",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 313,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "outputting solutions to related prompts that appear more frequently in the training set. To elicit this",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 324,
+ 505,
+ 338
+ ],
+ "score": 1.0,
+ "content": "failure, we start with prompts that apply a unary operation before a binary operation (unary-first), then",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 349
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 335,
+ 505,
+ 349
+ ],
+ "score": 1.0,
+ "content": "flip the order (binary-first). Programmers tend to apply unary operations first (e.g. when computing",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 346,
+ 505,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 346,
+ 505,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Euclidean distances or variances), so we conjecture that they appear more frequently on GitHub. We",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 357,
+ 505,
+ 370
+ ],
+ "score": 1.0,
+ "content": "first check that flipping the order of operations decreases accuracy. Next, to test if code generation",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 505,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 505,
+ 381
+ ],
+ "score": 1.0,
+ "content": "models instead outputs related prompts that occur more frequently in the training set, we measure",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 379,
+ 392,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 379,
+ 392,
+ 391
+ ],
+ "score": 1.0,
+ "content": "whether code generation models instead output the unary-first solution.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 22.5,
+ "bbox_fs": [
+ 105,
+ 302,
+ 505,
+ 391
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 395,
+ 505,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 394,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "We consider all 12 combinations of the binary operations sum, difference, and product, with unary",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 405,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "operations square, cube, quadruple, and square root. Focusing on Codex,5 we find that accuracy",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 417,
+ 505,
+ 429
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 417,
+ 151,
+ 429
+ ],
+ "score": 1.0,
+ "content": "drops from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 152,
+ 417,
+ 171,
+ 428
+ ],
+ "score": 0.87,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 171,
+ 417,
+ 182,
+ 429
+ ],
+ "score": 1.0,
+ "content": "to",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 182,
+ 417,
+ 201,
+ 428
+ ],
+ "score": 0.87,
+ "content": "17 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 201,
+ 417,
+ 505,
+ 429
+ ],
+ "score": 1.0,
+ "content": "when flipping the order from unary-first to binary-first. Among combinations",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 428,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 428,
+ 309,
+ 441
+ ],
+ "score": 1.0,
+ "content": "where flipping the order leads to error, we find that",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 309,
+ 428,
+ 329,
+ 439
+ ],
+ "score": 0.88,
+ "content": "7 5 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 428,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "of the binary-first outputs are the unary-first",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 439,
+ 281,
+ 452
+ ],
+ "score": 1.0,
+ "content": "solution. We exhibit one such error in Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 282,
+ 438,
+ 292,
+ 451
+ ],
+ "score": 0.46,
+ "content": "\\boxed { 5 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 292,
+ 439,
+ 505,
+ 452
+ ],
+ "score": 1.0,
+ "content": "when prompted to square the sum of its inputs, Codex",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 450,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "generates the correct function name (square_sum ), but reverses the order of operations. Our results",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 460,
+ 497,
+ 474
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 460,
+ 497,
+ 474
+ ],
+ "score": 1.0,
+ "content": "suggest that Codex can err by outputting solutions to related, frequent prompts in the training set.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 394,
+ 505,
+ 474
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 477,
+ 505,
+ 522
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 506,
+ 489
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 477,
+ 506,
+ 489
+ ],
+ "score": 1.0,
+ "content": "Control experiments. One worry is that the dip in performance is due the instructional nature of our",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 500
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 488,
+ 506,
+ 500
+ ],
+ "score": 1.0,
+ "content": "prompts. We rule this out by evaluating Codex on prompts where the docstring appears beneath the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 511
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 500,
+ 505,
+ 511
+ ],
+ "score": 1.0,
+ "content": "function signature and is a definition rather than command, to more closely mimic some functions on",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 509,
+ 487,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 509,
+ 487,
+ 524
+ ],
+ "score": 1.0,
+ "content": "GitHub. We obtain qualitatively similar results on these prompts, see Appendix A.4 for details.",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 477,
+ 506,
+ 524
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 533,
+ 284,
+ 544
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 285,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 285,
+ 546
+ ],
+ "score": 1.0,
+ "content": "3.3.4 Inspiration: Attribute substitution",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 551,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 506,
+ 564
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 550,
+ 506,
+ 564
+ ],
+ "score": 1.0,
+ "content": "Finally, we draw inspiration from attribute substitution: the human tendency to respond to a com-",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 562,
+ 506,
+ 574
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 562,
+ 506,
+ 574
+ ],
+ "score": 1.0,
+ "content": "plicated question using a simpler, related question [Kahneman and Frederick, 2002]. For example,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 104,
+ 573,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 573,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "a professor when asked how likely a candidate is to be tenured, may instead respond with how",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 585,
+ 253,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 585,
+ 253,
+ 595
+ ],
+ "score": 1.0,
+ "content": "impressive they found their job talk.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 40.5,
+ "bbox_fs": [
+ 104,
+ 550,
+ 506,
+ 595
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 600,
+ 505,
+ 666
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 600,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "Using attribute substitution as inspiration, we hypothesize that Codex may use simple-but-incorrect",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 611,
+ 506,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 611,
+ 506,
+ 624
+ ],
+ "score": 1.0,
+ "content": "heuristics to generate solutions. To elicit this failure, we add requests for conflicting function names",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 506,
+ 636
+ ],
+ "score": 1.0,
+ "content": "to MathEquation prompts. For example, in Figure 5 we prompt Codex to write a program that sums",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 633,
+ 505,
+ 645
+ ],
+ "score": 1.0,
+ "content": "its inputs called product_plus_2 . We first check that adding conflicting function names decreases",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 643,
+ 506,
+ 657
+ ],
+ "score": 1.0,
+ "content": "Codex’s functional accuracy. Next, to test if Codex uses simple-but-incorrect heuristics to generate",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 419,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 419,
+ 667
+ ],
+ "score": 1.0,
+ "content": "solutions, we check whether the generate solution matches the function name.",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 45.5,
+ "bbox_fs": [
+ 105,
+ 600,
+ 506,
+ 667
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 671,
+ 503,
+ 694
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 671,
+ 505,
+ 684
+ ],
+ "score": 1.0,
+ "content": "We evaluate Codex using 90 MathEquation prompts where the desired solution and requested",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 682,
+ 505,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 682,
+ 505,
+ 695
+ ],
+ "score": 1.0,
+ "content": "function name differ. To construct prompts, we begin with a prompt that Codex originally solves",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49.5,
+ "bbox_fs": [
+ 106,
+ 671,
+ 505,
+ 695
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "score": 0.971,
+ "html": "| Name location | Correct | Matches function name | Other error |
| No name | 100.0 | = | 0.0 |
| Docstring | 4.4 | 80.0 | 15.6 |
| Function signature | 4.4 | 70.0 | 25.6 |
| Name first | 4.6 | 51.7 | 43.7 |
",
+ "type": "table",
+ "image_path": "81dcce63968dbcc3abba5ddacec3c1af131342a4d2b0f1dc64d31ff5f93f5c32.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 89.33333333333333
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 167,
+ 89.33333333333333,
+ 445,
+ 109.66666666666666
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 167,
+ 109.66666666666666,
+ 445,
+ 130.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 132,
+ 506,
+ 177
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Table 2: Results of the attribute substitution experiments. We report accuracy when we do not request",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 143,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 143,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "a contradictory function name (no name), we request a function name in the docstring (docstring), in",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the function signature below the docstring (function signature), or above the docstring (name first).",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 456,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 456,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Overall, we find that Codex frequently generates solutions based on the function name.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 225
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "(sum, difference, or product), then append a request for a specific, contradictory function name (see",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 295,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 146,
+ 226
+ ],
+ "score": 1.0,
+ "content": "Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 212,
+ 166,
+ 226
+ ],
+ "score": 0.29,
+ "content": "\\underline { { \\vert \\mathbf { A . } 4 \\vert } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 213,
+ 295,
+ 226
+ ],
+ "score": 1.0,
+ "content": "for full implementation details).",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 228,
+ 506,
+ 295
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 276,
+ 242
+ ],
+ "score": 1.0,
+ "content": "We report our experimental results in Table",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 276,
+ 229,
+ 287,
+ 241
+ ],
+ "score": 0.49,
+ "content": "\\triangledown",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 228,
+ 506,
+ 242
+ ],
+ "score": 1.0,
+ "content": "When we request a conflicting function name, Codex’s",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 191,
+ 253
+ ],
+ "score": 1.0,
+ "content": "accuracy drops from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 241,
+ 216,
+ 251
+ ],
+ "score": 0.87,
+ "content": "100 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 240,
+ 247,
+ 253
+ ],
+ "score": 1.0,
+ "content": "to only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 241,
+ 294,
+ 251
+ ],
+ "score": 0.83,
+ "content": "4 . 4 \\% - 4 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 294,
+ 240,
+ 505,
+ 253
+ ],
+ "score": 1.0,
+ "content": ". This finding holds whether we request the function",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 263
+ ],
+ "score": 1.0,
+ "content": "name in the docstring, write it in the function signature below the docstring, or write the function",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 261,
+ 507,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 261,
+ 397,
+ 276
+ ],
+ "score": 1.0,
+ "content": "name over a simple description on the function. Moreover, for between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 262,
+ 417,
+ 273
+ ],
+ "score": 0.87,
+ "content": "52 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 261,
+ 435,
+ 276
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 262,
+ 456,
+ 273
+ ],
+ "score": 0.87,
+ "content": "80 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 261,
+ 507,
+ 276
+ ],
+ "score": 1.0,
+ "content": "of prompts,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 274,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 274,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "Codex responds with the function specified in the function name. Our results indicate that Codex can",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 285,
+ 368,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 285,
+ 368,
+ 297
+ ],
+ "score": 1.0,
+ "content": "err by using simple-but-incorrect heuristics to generate solutions.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 310,
+ 200,
+ 323
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 308,
+ 201,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 308,
+ 201,
+ 326
+ ],
+ "score": 1.0,
+ "content": "4 GPT-3 Results",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 505,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 332,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 332,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "In this section, we extend our study from Codex to GPT-3. To test GPT-3 for failure modes, we try",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "to faithfully reproduce and extend the anchoring experiment of Jacowitz and Kahneman [1995] and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 349,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 349,
+ 369
+ ],
+ "score": 1.0,
+ "content": "framing effect experiment of Tversky and Kahneman [1981].",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 386
+ ],
+ "score": 1.0,
+ "content": "Anchoring. As in Section 3.3.2 we study Section 3.3.2, we study anchoring: humans’ tendency",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 504,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 504,
+ 397
+ ],
+ "score": 1.0,
+ "content": "to insufficiently adjust their estimates away from an initial value [Tversky and Kahneman, 1974].",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 393,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 393,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "We largely replicate the anchoring study presented in Jacowitz and Kahneman [1995], but test the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 404,
+ 355,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 404,
+ 355,
+ 418
+ ],
+ "score": 1.0,
+ "content": "“davinci-001” version of OpenAI’s GPT-3 instead of humans.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 421,
+ 505,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 422,
+ 505,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 422,
+ 505,
+ 434
+ ],
+ "score": 1.0,
+ "content": "In their original experiment, Jacowitz and Kahneman asked students to estimate quantities such as the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 432,
+ 506,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 432,
+ 506,
+ 446
+ ],
+ "score": 1.0,
+ "content": "length of the Mississippi river in miles. They then asked new students to estimate the same quantities,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 457
+ ],
+ "score": 1.0,
+ "content": "but first gave them a upper or lower bound on the true answer (e.g. the Mississippi river is longer than",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "700 miles), which they call anchors. They find that students tend to underestimate the true quantity",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 497,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 497,
+ 479
+ ],
+ "score": 1.0,
+ "content": "when prompted with the lower anchor, and overestimate it when prompted with the upper anchor.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 481,
+ 505,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "We adapt the anchoring study from Jacowitz and Kahneman [1995] by finding the true answer for 14",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "of their 15 original questions6, then computing upper and lower anchors by increasing and decreasing",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 507,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 253,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the true answer by a fixed percentage",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 506,
+ 259,
+ 515
+ ],
+ "score": 0.68,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 503,
+ 319,
+ 516
+ ],
+ "score": 1.0,
+ "content": ". See Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 503,
+ 337,
+ 514
+ ],
+ "score": 0.82,
+ "content": "\\mathbf { B . l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 503,
+ 507,
+ 516
+ ],
+ "score": 1.0,
+ "content": "for a full list of questions and true answers.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 303,
+ 527
+ ],
+ "score": 1.0,
+ "content": "As an example, if the actual answer is 2000 and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 516,
+ 310,
+ 526
+ ],
+ "score": 0.81,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 515,
+ 320,
+ 527
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 515,
+ 340,
+ 525
+ ],
+ "score": 0.87,
+ "content": "5 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 515,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": ", the upper anchor is 3000 and the lower",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 438,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 438,
+ 538
+ ],
+ "score": 1.0,
+ "content": "anchor is 1000. We use this bound as an anchor, so that a typical prompt might be:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 173,
+ 571,
+ 427,
+ 583
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 570,
+ 428,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 570,
+ 428,
+ 584
+ ],
+ "score": 1.0,
+ "content": "What is the length of the Mississippi River (in miles)? Answer:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 589,
+ 505,
+ 667
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 602
+ ],
+ "score": 1.0,
+ "content": "To study anchoring in GPT-3, we measure how prepending the anchor changes GPT-3’s estimate.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "We categorize four potential changes: the estimate does not change, the estimate shifts towards the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 611,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "anchor, the estimate shifts away from the anchor, and the estimate is gibberish. We report the results",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 140,
+ 635
+ ],
+ "score": 1.0,
+ "content": "in Table",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 140,
+ 621,
+ 150,
+ 635
+ ],
+ "score": 0.7,
+ "content": "^ 3 .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 622,
+ 163,
+ 635
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 163,
+ 622,
+ 233,
+ 634
+ ],
+ "score": 0.85,
+ "content": "p \\in \\{ 2 0 \\% , 5 0 \\% \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": ". We find that GPT-3 routinely updates its estimate when an anchor",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "is prepended, and tends to shift the estimate towards the anchor. We also find that while GPT-3’s",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 329,
+ 656
+ ],
+ "score": 1.0,
+ "content": "updated estimate sometimes matches the anchor exactly",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 644,
+ 348,
+ 655
+ ],
+ "score": 0.86,
+ "content": "67 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "of the time), it also often lands between",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 401,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 401,
+ 667
+ ],
+ "score": 1.0,
+ "content": "the anchor and the original prediction, mirroring the behavior of humans.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 671,
+ 505,
+ 705
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "Our replication has a few limitations. Like the original study our sample size is small, we construct",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 682,
+ 507,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 682,
+ 361,
+ 696
+ ],
+ "score": 1.0,
+ "content": "prompts with templates, and many of the outputs—on average",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 682,
+ 381,
+ 693
+ ],
+ "score": 0.86,
+ "content": "41 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 682,
+ 507,
+ 696
+ ],
+ "score": 1.0,
+ "content": "—are gibberish. Nevertheless,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 694,
+ 400,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 694,
+ 400,
+ 705
+ ],
+ "score": 1.0,
+ "content": "our results suggest that GPT-3 incorporates the anchor during estimation.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "page_idx": 7,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 742,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 301,
+ 740,
+ 310,
+ 752
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 12,
+ "width": 9
+ }
+ ]
+ }
+ ]
+ },
+ {
+ "type": "discarded",
+ "bbox": [
+ 115,
+ 711,
+ 475,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 477,
+ 724
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 118,
+ 709,
+ 477,
+ 724
+ ],
+ "score": 1.0,
+ "content": "6 We omit a question asking for the number of bars in Berkeley, CA, since the answer is ambiguous",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "blocks": [
+ {
+ "type": "table_body",
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 130
+ ],
+ "score": 0.971,
+ "html": "| Name location | Correct | Matches function name | Other error |
| No name | 100.0 | = | 0.0 |
| Docstring | 4.4 | 80.0 | 15.6 |
| Function signature | 4.4 | 70.0 | 25.6 |
| Name first | 4.6 | 51.7 | 43.7 |
",
+ "type": "table",
+ "image_path": "81dcce63968dbcc3abba5ddacec3c1af131342a4d2b0f1dc64d31ff5f93f5c32.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 1,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 167,
+ 69,
+ 445,
+ 89.33333333333333
+ ],
+ "spans": [],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 167,
+ 89.33333333333333,
+ 445,
+ 109.66666666666666
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 167,
+ 109.66666666666666,
+ 445,
+ 130.0
+ ],
+ "spans": [],
+ "index": 2
+ }
+ ]
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 132,
+ 506,
+ 177
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 131,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "Table 2: Results of the attribute substitution experiments. We report accuracy when we do not request",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 143,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 143,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "a contradictory function name (no name), we request a function name in the docstring (docstring), in",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 154,
+ 507,
+ 167
+ ],
+ "score": 1.0,
+ "content": "the function signature below the docstring (function signature), or above the docstring (name first).",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 164,
+ 456,
+ 178
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 164,
+ 456,
+ 178
+ ],
+ "score": 1.0,
+ "content": "Overall, we find that Codex frequently generates solutions based on the function name.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 104,
+ 131,
+ 507,
+ 178
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 202,
+ 505,
+ 225
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 201,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "(sum, difference, or product), then append a request for a specific, contradictory function name (see",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 105,
+ 212,
+ 295,
+ 226
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 213,
+ 146,
+ 226
+ ],
+ "score": 1.0,
+ "content": "Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 212,
+ 166,
+ 226
+ ],
+ "score": 0.29,
+ "content": "\\underline { { \\vert \\mathbf { A . } 4 \\vert } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 166,
+ 213,
+ 295,
+ 226
+ ],
+ "score": 1.0,
+ "content": "for full implementation details).",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7.5,
+ "bbox_fs": [
+ 105,
+ 201,
+ 506,
+ 226
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 228,
+ 506,
+ 295
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 506,
+ 242
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 228,
+ 276,
+ 242
+ ],
+ "score": 1.0,
+ "content": "We report our experimental results in Table",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 276,
+ 229,
+ 287,
+ 241
+ ],
+ "score": 0.49,
+ "content": "\\triangledown",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 287,
+ 228,
+ 506,
+ 242
+ ],
+ "score": 1.0,
+ "content": "When we request a conflicting function name, Codex’s",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 191,
+ 253
+ ],
+ "score": 1.0,
+ "content": "accuracy drops from",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 192,
+ 241,
+ 216,
+ 251
+ ],
+ "score": 0.87,
+ "content": "100 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 216,
+ 240,
+ 247,
+ 253
+ ],
+ "score": 1.0,
+ "content": "to only",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 241,
+ 294,
+ 251
+ ],
+ "score": 0.83,
+ "content": "4 . 4 \\% - 4 . 6 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 294,
+ 240,
+ 505,
+ 253
+ ],
+ "score": 1.0,
+ "content": ". This finding holds whether we request the function",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 263
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 252,
+ 505,
+ 263
+ ],
+ "score": 1.0,
+ "content": "name in the docstring, write it in the function signature below the docstring, or write the function",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 104,
+ 261,
+ 507,
+ 276
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 261,
+ 397,
+ 276
+ ],
+ "score": 1.0,
+ "content": "name over a simple description on the function. Moreover, for between",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 397,
+ 262,
+ 417,
+ 273
+ ],
+ "score": 0.87,
+ "content": "52 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 418,
+ 261,
+ 435,
+ 276
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 436,
+ 262,
+ 456,
+ 273
+ ],
+ "score": 0.87,
+ "content": "80 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 456,
+ 261,
+ 507,
+ 276
+ ],
+ "score": 1.0,
+ "content": "of prompts,",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 106,
+ 274,
+ 505,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 274,
+ 505,
+ 285
+ ],
+ "score": 1.0,
+ "content": "Codex responds with the function specified in the function name. Our results indicate that Codex can",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 106,
+ 285,
+ 368,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 285,
+ 368,
+ 297
+ ],
+ "score": 1.0,
+ "content": "err by using simple-but-incorrect heuristics to generate solutions.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 11.5,
+ "bbox_fs": [
+ 104,
+ 228,
+ 507,
+ 297
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 310,
+ 200,
+ 323
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 308,
+ 201,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 308,
+ 201,
+ 326
+ ],
+ "score": 1.0,
+ "content": "4 GPT-3 Results",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 505,
+ 368
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 332,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 332,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "In this section, we extend our study from Codex to GPT-3. To test GPT-3 for failure modes, we try",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 344,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "to faithfully reproduce and extend the anchoring experiment of Jacowitz and Kahneman [1995] and",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 105,
+ 356,
+ 349,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 356,
+ 349,
+ 369
+ ],
+ "score": 1.0,
+ "content": "framing effect experiment of Tversky and Kahneman [1981].",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 104,
+ 332,
+ 506,
+ 369
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 372,
+ 505,
+ 416
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 386
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 371,
+ 505,
+ 386
+ ],
+ "score": 1.0,
+ "content": "Anchoring. As in Section 3.3.2 we study Section 3.3.2, we study anchoring: humans’ tendency",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 383,
+ 504,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 383,
+ 504,
+ 397
+ ],
+ "score": 1.0,
+ "content": "to insufficiently adjust their estimates away from an initial value [Tversky and Kahneman, 1974].",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 104,
+ 393,
+ 505,
+ 407
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 393,
+ 505,
+ 407
+ ],
+ "score": 1.0,
+ "content": "We largely replicate the anchoring study presented in Jacowitz and Kahneman [1995], but test the",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 104,
+ 404,
+ 355,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 404,
+ 355,
+ 418
+ ],
+ "score": 1.0,
+ "content": "“davinci-001” version of OpenAI’s GPT-3 instead of humans.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 20.5,
+ "bbox_fs": [
+ 104,
+ 371,
+ 505,
+ 418
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 421,
+ 505,
+ 477
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 422,
+ 505,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 422,
+ 505,
+ 434
+ ],
+ "score": 1.0,
+ "content": "In their original experiment, Jacowitz and Kahneman asked students to estimate quantities such as the",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 106,
+ 432,
+ 506,
+ 446
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 432,
+ 506,
+ 446
+ ],
+ "score": 1.0,
+ "content": "length of the Mississippi river in miles. They then asked new students to estimate the same quantities,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 442,
+ 506,
+ 457
+ ],
+ "score": 1.0,
+ "content": "but first gave them a upper or lower bound on the true answer (e.g. the Mississippi river is longer than",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 468
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 453,
+ 505,
+ 468
+ ],
+ "score": 1.0,
+ "content": "700 miles), which they call anchors. They find that students tend to underestimate the true quantity",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 465,
+ 497,
+ 479
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 465,
+ 497,
+ 479
+ ],
+ "score": 1.0,
+ "content": "when prompted with the lower anchor, and overestimate it when prompted with the upper anchor.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 25,
+ "bbox_fs": [
+ 105,
+ 422,
+ 506,
+ 479
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 481,
+ 505,
+ 537
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 482,
+ 505,
+ 494
+ ],
+ "score": 1.0,
+ "content": "We adapt the anchoring study from Jacowitz and Kahneman [1995] by finding the true answer for 14",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 490,
+ 506,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 490,
+ 506,
+ 506
+ ],
+ "score": 1.0,
+ "content": "of their 15 original questions6, then computing upper and lower anchors by increasing and decreasing",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 503,
+ 507,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 503,
+ 253,
+ 516
+ ],
+ "score": 1.0,
+ "content": "the true answer by a fixed percentage",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 253,
+ 506,
+ 259,
+ 515
+ ],
+ "score": 0.68,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 259,
+ 503,
+ 319,
+ 516
+ ],
+ "score": 1.0,
+ "content": ". See Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 503,
+ 337,
+ 514
+ ],
+ "score": 0.82,
+ "content": "\\mathbf { B . l }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 338,
+ 503,
+ 507,
+ 516
+ ],
+ "score": 1.0,
+ "content": "for a full list of questions and true answers.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 303,
+ 527
+ ],
+ "score": 1.0,
+ "content": "As an example, if the actual answer is 2000 and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 303,
+ 516,
+ 310,
+ 526
+ ],
+ "score": 0.81,
+ "content": "p",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 310,
+ 515,
+ 320,
+ 527
+ ],
+ "score": 1.0,
+ "content": "is",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 320,
+ 515,
+ 340,
+ 525
+ ],
+ "score": 0.87,
+ "content": "5 0 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 340,
+ 515,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": ", the upper anchor is 3000 and the lower",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 524,
+ 438,
+ 538
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 524,
+ 438,
+ 538
+ ],
+ "score": 1.0,
+ "content": "anchor is 1000. We use this bound as an anchor, so that a typical prompt might be:",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 482,
+ 507,
+ 538
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 173,
+ 571,
+ 427,
+ 583
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 172,
+ 570,
+ 428,
+ 584
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 172,
+ 570,
+ 428,
+ 584
+ ],
+ "score": 1.0,
+ "content": "What is the length of the Mississippi River (in miles)? Answer:",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 33,
+ "bbox_fs": [
+ 172,
+ 570,
+ 428,
+ 584
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 589,
+ 505,
+ 667
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 602
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 588,
+ 506,
+ 602
+ ],
+ "score": 1.0,
+ "content": "To study anchoring in GPT-3, we measure how prepending the anchor changes GPT-3’s estimate.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 600,
+ 505,
+ 613
+ ],
+ "score": 1.0,
+ "content": "We categorize four potential changes: the estimate does not change, the estimate shifts towards the",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 106,
+ 611,
+ 505,
+ 624
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 611,
+ 505,
+ 624
+ ],
+ "score": 1.0,
+ "content": "anchor, the estimate shifts away from the anchor, and the estimate is gibberish. We report the results",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 621,
+ 505,
+ 635
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 622,
+ 140,
+ 635
+ ],
+ "score": 1.0,
+ "content": "in Table",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 140,
+ 621,
+ 150,
+ 635
+ ],
+ "score": 0.7,
+ "content": "^ 3 .",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 150,
+ 622,
+ 163,
+ 635
+ ],
+ "score": 1.0,
+ "content": "for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 163,
+ 622,
+ 233,
+ 634
+ ],
+ "score": 0.85,
+ "content": "p \\in \\{ 2 0 \\% , 5 0 \\% \\}",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 233,
+ 622,
+ 505,
+ 635
+ ],
+ "score": 1.0,
+ "content": ". We find that GPT-3 routinely updates its estimate when an anchor",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 104,
+ 632,
+ 506,
+ 645
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 632,
+ 506,
+ 645
+ ],
+ "score": 1.0,
+ "content": "is prepended, and tends to shift the estimate towards the anchor. We also find that while GPT-3’s",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 105,
+ 644,
+ 506,
+ 656
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 644,
+ 329,
+ 656
+ ],
+ "score": 1.0,
+ "content": "updated estimate sometimes matches the anchor exactly",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 329,
+ 644,
+ 348,
+ 655
+ ],
+ "score": 0.86,
+ "content": "67 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 349,
+ 644,
+ 506,
+ 656
+ ],
+ "score": 1.0,
+ "content": "of the time), it also often lands between",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 655,
+ 401,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 655,
+ 401,
+ 667
+ ],
+ "score": 1.0,
+ "content": "the anchor and the original prediction, mirroring the behavior of humans.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 37,
+ "bbox_fs": [
+ 104,
+ 588,
+ 506,
+ 667
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 671,
+ 505,
+ 705
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 506,
+ 684
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 670,
+ 506,
+ 684
+ ],
+ "score": 1.0,
+ "content": "Our replication has a few limitations. Like the original study our sample size is small, we construct",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 106,
+ 682,
+ 507,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 682,
+ 361,
+ 696
+ ],
+ "score": 1.0,
+ "content": "prompts with templates, and many of the outputs—on average",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 362,
+ 682,
+ 381,
+ 693
+ ],
+ "score": 0.86,
+ "content": "41 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 382,
+ 682,
+ 507,
+ 696
+ ],
+ "score": 1.0,
+ "content": "—are gibberish. Nevertheless,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 694,
+ 400,
+ 705
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 694,
+ 400,
+ 705
+ ],
+ "score": 1.0,
+ "content": "our results suggest that GPT-3 incorporates the anchor during estimation.",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 106,
+ 670,
+ 507,
+ 705
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 258,
+ 71,
+ 381,
+ 80
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 258,
+ 70,
+ 381,
+ 82
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 258,
+ 70,
+ 381,
+ 82
+ ],
+ "score": 1.0,
+ "content": "Anchor change to model output",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "score": 0.94,
+ "html": "| p | No change | Towards anchor | Away from anchor | Gibberish |
| 20% | 10.7 | 28.6 | 10.7 | 50.0 |
| 50% | 14.3 | 42.9 | 10.7 | 32.1 |
",
+ "type": "table",
+ "image_path": "ca122c6a93f0d5e07759a579d77303538770f464f3628eed3bba4e6c8f66b75d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 93.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 169,
+ 93.0,
+ 441,
+ 106.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 169,
+ 106.0,
+ 441,
+ 119.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ }
+ ],
+ "index": 1.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 167
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 463,
+ 134
+ ],
+ "score": 1.0,
+ "content": "Table 3: Results of the adaptation of the anchoring study from Jacowitz and Kahneman",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 121,
+ 491,
+ 135
+ ],
+ "score": 1.0,
+ "content": "[1995]",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 123,
+ 506,
+ 133
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 257,
+ 145
+ ],
+ "score": 1.0,
+ "content": "GPT-3. We consider anchors that are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 133,
+ 278,
+ 144
+ ],
+ "score": 0.88,
+ "content": "20 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 132,
+ 296,
+ 145
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 133,
+ 316,
+ 144
+ ],
+ "score": 0.88,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "increases and decreases from the ground truth",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 158
+ ],
+ "score": 1.0,
+ "content": "answer, and measure how often GPT-3’s revised prediction does not change, shifts towards / away",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 413,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 413,
+ 167
+ ],
+ "score": 1.0,
+ "content": "from the anchor, or is gibberish, aggregated across lower and upper anchors.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 110,
+ 186,
+ 340,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 187,
+ 124,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 187,
+ 124,
+ 192
+ ],
+ "score": 0.554,
+ "content": "un",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 110,
+ 195,
+ 340,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 195,
+ 340,
+ 204
+ ],
+ "score": 1.0,
+ "content": "Delete all files in an inputted directory that import all",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 110,
+ 201,
+ 280,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 201,
+ 280,
+ 212
+ ],
+ "score": 1.0,
+ "content": "of statsmodels, plotly, seaborn, and scipy",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 110,
+ 210,
+ 125,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 210,
+ 125,
+ 218
+ ],
+ "score": 1.0,
+ "content": "“””",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 9.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 111,
+ 221,
+ 348,
+ 284
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 110,
+ 219,
+ 149,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 219,
+ 149,
+ 228
+ ],
+ "score": 1.0,
+ "content": "import os",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 111,
+ 227,
+ 149,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 227,
+ 149,
+ 235
+ ],
+ "score": 1.0,
+ "content": "import re",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 110,
+ 234,
+ 216,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 234,
+ 216,
+ 244
+ ],
+ "score": 1.0,
+ "content": "def delete_all(directory):",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 124,
+ 243,
+ 293,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 243,
+ 293,
+ 252
+ ],
+ "score": 1.0,
+ "content": "for root, dir, file in os.walk(directory):",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 137,
+ 250,
+ 200,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 137,
+ 250,
+ 200,
+ 259
+ ],
+ "score": 1.0,
+ "content": "for f in files:",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 150,
+ 259,
+ 328,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 150,
+ 259,
+ 328,
+ 268
+ ],
+ "score": 1.0,
+ "content": "with open(os.path.join(root, f), ‘r’) as fp:",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 163,
+ 266,
+ 349,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 163,
+ 266,
+ 349,
+ 277
+ ],
+ "score": 1.0,
+ "content": "if re.search(“import statsmodels”, fp.read()):",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 177,
+ 275,
+ 307,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 177,
+ 275,
+ 307,
+ 285
+ ],
+ "score": 1.0,
+ "content": "os.remove(os.path.join(root, f))",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "score": 0.964,
+ "type": "image",
+ "image_path": "71ff0596f1e01e59c72b5c20ada600f5859e331d42315f4d75e7a8346414e74a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 234.5
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 354,
+ 234.5,
+ 501,
+ 282.0
+ ],
+ "spans": [],
+ "index": 21
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 289,
+ 505,
+ 355
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 289,
+ 506,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 289,
+ 506,
+ 302
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Left. Example where Codex incorrectly deletes files. We prompt Codex to delete files",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "containing all of statsmodels, plotly, seaborn, and scipy. Codex correctly iterates through all files",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 311,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "in the inputted directory (blue), but then incorrectly deletes all files containing statsmodels (red), as",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 505,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 505,
+ 334
+ ],
+ "score": 1.0,
+ "content": "attribute substitution suggests. Right. Plot describing the errors Codex makes as a function of the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "number of packages. We find that Codex often incorrectly deletes files if they contain any of the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 104,
+ 342,
+ 483,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 342,
+ 483,
+ 357
+ ],
+ "score": 1.0,
+ "content": "listed packages, and relies more on just the first package as the number of packages increases.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24.5
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 433
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 378,
+ 505,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 378,
+ 238,
+ 391
+ ],
+ "score": 1.0,
+ "content": "Framing effect. As in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 378,
+ 264,
+ 391
+ ],
+ "score": 0.82,
+ "content": "3 . 3 . 1 ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 378,
+ 505,
+ 391
+ ],
+ "score": 1.0,
+ "content": "we study the framing effect: predictable shifts in human",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 389,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "responses when the same problem is framed in different ways. We largely replicate the framing exper-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "iment presented in Tversky and Kahneman [1981]: we compare GPT-3’s responses to two equivalent",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 506,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 506,
+ 424
+ ],
+ "score": 1.0,
+ "content": "decisions: choosing to either deterministically save (or let die) some fraction of a population, or to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 315,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 315,
+ 434
+ ],
+ "score": 1.0,
+ "content": "probabilistically save (let die) the whole population.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 526
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "We measure the rate at which GPT-3 chooses the probabilistic option across different population sizes",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 310,
+ 462
+ ],
+ "score": 1.0,
+ "content": "and different fractions / probabilities. See Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 449,
+ 329,
+ 461
+ ],
+ "score": 0.5,
+ "content": "\\boxed { \\mathbf { B } . 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 450,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "for full results. When using the probability",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "in the original study, GPT-3 qualitatively mirrors humans: it chooses the probabilistic option far",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "more frequently under the “not save” framing than under the “save framing”. However, for higher",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "probabilities, GPT-3 consistently chooses the probabilistic option for both framings; we conjecture",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 506
+ ],
+ "score": 1.0,
+ "content": "that humans could exhibit similar behavior in this regime, since the probabilistic option is more",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "certain. Overall, our results suggest that GPT-3 selects different options based on the framing, and",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 488,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 488,
+ 527
+ ],
+ "score": 1.0,
+ "content": "could be a test-bed to identify qualitative human behaviors without running full human studies.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 543,
+ 229,
+ 557
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 542,
+ 231,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 542,
+ 231,
+ 560
+ ],
+ "score": 1.0,
+ "content": "5 High-Impact Errors",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 568,
+ 505,
+ 602
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "We have shown how our framework helps us elicit failures of large language models. In this section,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "we use our framework to construct cases where Codex makes high-impact errors: harmful errors that",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 470,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 470,
+ 604
+ ],
+ "score": 1.0,
+ "content": "are hard to undo. Specifically, we construct prompts where Codex incorrectly deletes files.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 696
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 159,
+ 619
+ ],
+ "score": 1.0,
+ "content": "As in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 606,
+ 183,
+ 619
+ ],
+ "score": 0.34,
+ "content": "3 . 3 . 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 607,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "we draw inspiration from attribute substitution: the tendency of humans to respond",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "to a complex question with a simpler, related question. Using attribute substitution as motivation, we",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "hypothesize that Codex may simplify complex expressions such as conjunctions. Instead of checking",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "all components of a conjunction at once, it might “give up” and consider subsets of the components",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 227,
+ 663
+ ],
+ "score": 1.0,
+ "content": "individually (e.g. checking for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 651,
+ 236,
+ 660
+ ],
+ "score": 0.77,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 651,
+ 248,
+ 663
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 651,
+ 275,
+ 661
+ ],
+ "score": 0.89,
+ "content": "A \\lor B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 651,
+ 317,
+ 663
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 651,
+ 345,
+ 661
+ ],
+ "score": 0.87,
+ "content": "A \\land B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 651,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "). To elicit this failure, we prompt Codex",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 662,
+ 380,
+ 674
+ ],
+ "score": 1.0,
+ "content": "to delete files containing specific sets of package imports; see Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 662,
+ 389,
+ 674
+ ],
+ "score": 0.69,
+ "content": "6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 390,
+ 662,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "for an example. We measure",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 672,
+ 506,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 672,
+ 506,
+ 685
+ ],
+ "score": 1.0,
+ "content": "how often Codex generates a simpler output that erroneously deletes files, as well as how often it",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 683,
+ 376,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 278,
+ 695
+ ],
+ "score": 1.0,
+ "content": "produces the correct output. See Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 683,
+ 289,
+ 696
+ ],
+ "score": 0.78,
+ "content": "\\boxed { \\mathbf { C } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 289,
+ 684,
+ 376,
+ 695
+ ],
+ "score": 1.0,
+ "content": "for additional details.",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 48.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 700,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "We test for two types of simpler outputs: code deleting all files containing first package in the set",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 125,
+ 723
+ ],
+ "score": 1.0,
+ "content": "(i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 126,
+ 711,
+ 134,
+ 721
+ ],
+ "score": 0.69,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 710,
+ 177,
+ 723
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 711,
+ 206,
+ 721
+ ],
+ "score": 0.87,
+ "content": "A \\land B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 710,
+ 475,
+ 723
+ ],
+ "score": 1.0,
+ "content": "), and code deleting all files containing any package in the set (i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 476,
+ 711,
+ 504,
+ 721
+ ],
+ "score": 0.81,
+ "content": "A \\lor B",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 54
+ }
+ ],
+ "index": 53.5
+ }
+ ],
+ "page_idx": 8,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 302,
+ 741,
+ 308,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 302,
+ 741,
+ 309,
+ 752
+ ],
+ "score": 1.0,
+ "content": "9",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "table",
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "blocks": [
+ {
+ "type": "table_caption",
+ "bbox": [
+ 258,
+ 71,
+ 381,
+ 80
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 258,
+ 70,
+ 381,
+ 82
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 258,
+ 70,
+ 381,
+ 82
+ ],
+ "score": 1.0,
+ "content": "Anchor change to model output",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "table_body",
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 119
+ ],
+ "score": 0.94,
+ "html": "| p | No change | Towards anchor | Away from anchor | Gibberish |
| 20% | 10.7 | 28.6 | 10.7 | 50.0 |
| 50% | 14.3 | 42.9 | 10.7 | 32.1 |
",
+ "type": "table",
+ "image_path": "ca122c6a93f0d5e07759a579d77303538770f464f3628eed3bba4e6c8f66b75d.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 2,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 169,
+ 80,
+ 441,
+ 93.0
+ ],
+ "spans": [],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 169,
+ 93.0,
+ 441,
+ 106.0
+ ],
+ "spans": [],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 169,
+ 106.0,
+ 441,
+ 119.0
+ ],
+ "spans": [],
+ "index": 3
+ }
+ ]
+ }
+ ],
+ "index": 1.0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 122,
+ 505,
+ 167
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 121,
+ 463,
+ 134
+ ],
+ "score": 1.0,
+ "content": "Table 3: Results of the adaptation of the anchoring study from Jacowitz and Kahneman",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 465,
+ 121,
+ 491,
+ 135
+ ],
+ "score": 1.0,
+ "content": "[1995]",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 492,
+ 123,
+ 506,
+ 133
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 505,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 257,
+ 145
+ ],
+ "score": 1.0,
+ "content": "GPT-3. We consider anchors that are",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 258,
+ 133,
+ 278,
+ 144
+ ],
+ "score": 0.88,
+ "content": "20 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 278,
+ 132,
+ 296,
+ 145
+ ],
+ "score": 1.0,
+ "content": "and",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 296,
+ 133,
+ 316,
+ 144
+ ],
+ "score": 0.88,
+ "content": "50 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 316,
+ 132,
+ 505,
+ 145
+ ],
+ "score": 1.0,
+ "content": "increases and decreases from the ground truth",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 143,
+ 506,
+ 158
+ ],
+ "score": 1.0,
+ "content": "answer, and measure how often GPT-3’s revised prediction does not change, shifts towards / away",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 413,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 413,
+ 167
+ ],
+ "score": 1.0,
+ "content": "from the anchor, or is gibberish, aggregated across lower and upper anchors.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 5.5,
+ "bbox_fs": [
+ 105,
+ 121,
+ 506,
+ 167
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 110,
+ 186,
+ 340,
+ 217
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 111,
+ 187,
+ 124,
+ 192
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 187,
+ 124,
+ 192
+ ],
+ "score": 0.554,
+ "content": "un",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 195,
+ 340,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 195,
+ 340,
+ 204
+ ],
+ "score": 1.0,
+ "content": "Delete all files in an inputted directory that import all",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 201,
+ 280,
+ 212
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 201,
+ 280,
+ 212
+ ],
+ "score": 1.0,
+ "content": "of statsmodels, plotly, seaborn, and scipy",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 210,
+ 125,
+ 218
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 210,
+ 125,
+ 218
+ ],
+ "score": 1.0,
+ "content": "“””",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 219,
+ 149,
+ 228
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 219,
+ 149,
+ 228
+ ],
+ "score": 1.0,
+ "content": "import os",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 111,
+ 227,
+ 149,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 111,
+ 227,
+ 149,
+ 235
+ ],
+ "score": 1.0,
+ "content": "import re",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 110,
+ 234,
+ 216,
+ 244
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 110,
+ 234,
+ 216,
+ 244
+ ],
+ "score": 1.0,
+ "content": "def delete_all(directory):",
+ "type": "text"
+ }
+ ],
+ "index": 15,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 124,
+ 243,
+ 293,
+ 252
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 124,
+ 243,
+ 293,
+ 252
+ ],
+ "score": 1.0,
+ "content": "for root, dir, file in os.walk(directory):",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 137,
+ 250,
+ 200,
+ 259
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 137,
+ 250,
+ 200,
+ 259
+ ],
+ "score": 1.0,
+ "content": "for f in files:",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 150,
+ 259,
+ 328,
+ 268
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 150,
+ 259,
+ 328,
+ 268
+ ],
+ "score": 1.0,
+ "content": "with open(os.path.join(root, f), ‘r’) as fp:",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 163,
+ 266,
+ 349,
+ 277
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 163,
+ 266,
+ 349,
+ 277
+ ],
+ "score": 1.0,
+ "content": "if re.search(“import statsmodels”, fp.read()):",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 177,
+ 275,
+ 307,
+ 285
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 177,
+ 275,
+ 307,
+ 285
+ ],
+ "score": 1.0,
+ "content": "os.remove(os.path.join(root, f))",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 9.5,
+ "bbox_fs": [
+ 110,
+ 187,
+ 340,
+ 218
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 111,
+ 221,
+ 348,
+ 284
+ ],
+ "lines": [],
+ "index": 16.5,
+ "bbox_fs": [
+ 110,
+ 219,
+ 349,
+ 285
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "image",
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "blocks": [
+ {
+ "type": "image_body",
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 282
+ ],
+ "score": 0.964,
+ "type": "image",
+ "image_path": "71ff0596f1e01e59c72b5c20ada600f5859e331d42315f4d75e7a8346414e74a.jpg"
+ }
+ ]
+ }
+ ],
+ "index": 16.5,
+ "virtual_lines": [
+ {
+ "bbox": [
+ 354,
+ 187,
+ 501,
+ 234.5
+ ],
+ "spans": [],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 354,
+ 234.5,
+ 501,
+ 282.0
+ ],
+ "spans": [],
+ "index": 21
+ }
+ ]
+ },
+ {
+ "type": "image_caption",
+ "bbox": [
+ 106,
+ 289,
+ 505,
+ 355
+ ],
+ "group_id": 0,
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 289,
+ 506,
+ 302
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 289,
+ 506,
+ 302
+ ],
+ "score": 1.0,
+ "content": "Figure 6: Left. Example where Codex incorrectly deletes files. We prompt Codex to delete files",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 300,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "containing all of statsmodels, plotly, seaborn, and scipy. Codex correctly iterates through all files",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 311,
+ 505,
+ 324
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 311,
+ 505,
+ 324
+ ],
+ "score": 1.0,
+ "content": "in the inputted directory (blue), but then incorrectly deletes all files containing statsmodels (red), as",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 105,
+ 322,
+ 505,
+ 334
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 322,
+ 505,
+ 334
+ ],
+ "score": 1.0,
+ "content": "attribute substitution suggests. Right. Plot describing the errors Codex makes as a function of the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 333,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "number of packages. We find that Codex often incorrectly deletes files if they contain any of the",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 104,
+ 342,
+ 483,
+ 357
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 342,
+ 483,
+ 357
+ ],
+ "score": 1.0,
+ "content": "listed packages, and relies more on just the first package as the number of packages increases.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 24.5
+ }
+ ],
+ "index": 20.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 377,
+ 505,
+ 433
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 378,
+ 505,
+ 391
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 378,
+ 238,
+ 391
+ ],
+ "score": 1.0,
+ "content": "Framing effect. As in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 239,
+ 378,
+ 264,
+ 391
+ ],
+ "score": 0.82,
+ "content": "3 . 3 . 1 ,",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 264,
+ 378,
+ 505,
+ 391
+ ],
+ "score": 1.0,
+ "content": "we study the framing effect: predictable shifts in human",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 106,
+ 389,
+ 506,
+ 401
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 389,
+ 506,
+ 401
+ ],
+ "score": 1.0,
+ "content": "responses when the same problem is framed in different ways. We largely replicate the framing exper-",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 413
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 399,
+ 506,
+ 413
+ ],
+ "score": 1.0,
+ "content": "iment presented in Tversky and Kahneman [1981]: we compare GPT-3’s responses to two equivalent",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 410,
+ 506,
+ 424
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 410,
+ 506,
+ 424
+ ],
+ "score": 1.0,
+ "content": "decisions: choosing to either deterministically save (or let die) some fraction of a population, or to",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 105,
+ 422,
+ 315,
+ 434
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 422,
+ 315,
+ 434
+ ],
+ "score": 1.0,
+ "content": "probabilistically save (let die) the whole population.",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 378,
+ 506,
+ 434
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 526
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 438,
+ 505,
+ 450
+ ],
+ "score": 1.0,
+ "content": "We measure the rate at which GPT-3 chooses the probabilistic option across different population sizes",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 449,
+ 505,
+ 462
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 450,
+ 310,
+ 462
+ ],
+ "score": 1.0,
+ "content": "and different fractions / probabilities. See Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 310,
+ 449,
+ 329,
+ 461
+ ],
+ "score": 0.5,
+ "content": "\\boxed { \\mathbf { B } . 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 329,
+ 450,
+ 505,
+ 462
+ ],
+ "score": 1.0,
+ "content": "for full results. When using the probability",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 460,
+ 505,
+ 473
+ ],
+ "score": 1.0,
+ "content": "in the original study, GPT-3 qualitatively mirrors humans: it chooses the probabilistic option far",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 483
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 471,
+ 505,
+ 483
+ ],
+ "score": 1.0,
+ "content": "more frequently under the “not save” framing than under the “save framing”. However, for higher",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 482,
+ 505,
+ 495
+ ],
+ "score": 1.0,
+ "content": "probabilities, GPT-3 consistently chooses the probabilistic option for both framings; we conjecture",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 506
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 492,
+ 505,
+ 506
+ ],
+ "score": 1.0,
+ "content": "that humans could exhibit similar behavior in this regime, since the probabilistic option is more",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 504,
+ 505,
+ 516
+ ],
+ "score": 1.0,
+ "content": "certain. Overall, our results suggest that GPT-3 selects different options based on the framing, and",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 106,
+ 515,
+ 488,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 515,
+ 488,
+ 527
+ ],
+ "score": 1.0,
+ "content": "could be a test-bed to identify qualitative human behaviors without running full human studies.",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 36.5,
+ "bbox_fs": [
+ 105,
+ 438,
+ 505,
+ 527
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 543,
+ 229,
+ 557
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 542,
+ 231,
+ 560
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 542,
+ 231,
+ 560
+ ],
+ "score": 1.0,
+ "content": "5 High-Impact Errors",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 568,
+ 505,
+ 602
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 582
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 568,
+ 505,
+ 582
+ ],
+ "score": 1.0,
+ "content": "We have shown how our framework helps us elicit failures of large language models. In this section,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 580,
+ 505,
+ 592
+ ],
+ "score": 1.0,
+ "content": "we use our framework to construct cases where Codex makes high-impact errors: harmful errors that",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 591,
+ 470,
+ 604
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 591,
+ 470,
+ 604
+ ],
+ "score": 1.0,
+ "content": "are hard to undo. Specifically, we construct prompts where Codex incorrectly deletes files.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43,
+ "bbox_fs": [
+ 105,
+ 568,
+ 505,
+ 604
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 696
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 606,
+ 505,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 607,
+ 159,
+ 619
+ ],
+ "score": 1.0,
+ "content": "As in Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 159,
+ 606,
+ 183,
+ 619
+ ],
+ "score": 0.34,
+ "content": "3 . 3 . 4",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 183,
+ 607,
+ 505,
+ 619
+ ],
+ "score": 1.0,
+ "content": "we draw inspiration from attribute substitution: the tendency of humans to respond",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 618,
+ 505,
+ 630
+ ],
+ "score": 1.0,
+ "content": "to a complex question with a simpler, related question. Using attribute substitution as motivation, we",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 628,
+ 505,
+ 642
+ ],
+ "score": 1.0,
+ "content": "hypothesize that Codex may simplify complex expressions such as conjunctions. Instead of checking",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 653
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 639,
+ 505,
+ 653
+ ],
+ "score": 1.0,
+ "content": "all components of a conjunction at once, it might “give up” and consider subsets of the components",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 106,
+ 651,
+ 505,
+ 663
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 651,
+ 227,
+ 663
+ ],
+ "score": 1.0,
+ "content": "individually (e.g. checking for",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 227,
+ 651,
+ 236,
+ 660
+ ],
+ "score": 0.77,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 236,
+ 651,
+ 248,
+ 663
+ ],
+ "score": 1.0,
+ "content": "or",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 248,
+ 651,
+ 275,
+ 661
+ ],
+ "score": 0.89,
+ "content": "A \\lor B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 275,
+ 651,
+ 317,
+ 663
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 317,
+ 651,
+ 345,
+ 661
+ ],
+ "score": 0.87,
+ "content": "A \\land B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 345,
+ 651,
+ 505,
+ 663
+ ],
+ "score": 1.0,
+ "content": "). To elicit this failure, we prompt Codex",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 106,
+ 662,
+ 505,
+ 674
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 662,
+ 380,
+ 674
+ ],
+ "score": 1.0,
+ "content": "to delete files containing specific sets of package imports; see Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 380,
+ 662,
+ 389,
+ 674
+ ],
+ "score": 0.69,
+ "content": "6",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 390,
+ 662,
+ 505,
+ 674
+ ],
+ "score": 1.0,
+ "content": "for an example. We measure",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 105,
+ 672,
+ 506,
+ 685
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 672,
+ 506,
+ 685
+ ],
+ "score": 1.0,
+ "content": "how often Codex generates a simpler output that erroneously deletes files, as well as how often it",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ },
+ {
+ "bbox": [
+ 105,
+ 683,
+ 376,
+ 696
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 684,
+ 278,
+ 695
+ ],
+ "score": 1.0,
+ "content": "produces the correct output. See Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 278,
+ 683,
+ 289,
+ 696
+ ],
+ "score": 0.78,
+ "content": "\\boxed { \\mathbf { C } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 289,
+ 684,
+ 376,
+ 695
+ ],
+ "score": 1.0,
+ "content": "for additional details.",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ }
+ ],
+ "index": 48.5,
+ "bbox_fs": [
+ 105,
+ 606,
+ 506,
+ 696
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 700,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "We test for two types of simpler outputs: code deleting all files containing first package in the set",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 504,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 125,
+ 723
+ ],
+ "score": 1.0,
+ "content": "(i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 126,
+ 711,
+ 134,
+ 721
+ ],
+ "score": 0.69,
+ "content": "A",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 710,
+ 177,
+ 723
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 178,
+ 711,
+ 206,
+ 721
+ ],
+ "score": 0.87,
+ "content": "A \\land B",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 207,
+ 710,
+ 475,
+ 723
+ ],
+ "score": 1.0,
+ "content": "), and code deleting all files containing any package in the set (i.e.",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 476,
+ 711,
+ 504,
+ 721
+ ],
+ "score": 0.81,
+ "content": "A \\lor B",
+ "type": "inline_equation"
+ }
+ ],
+ "index": 54
+ },
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 147,
+ 85
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 148,
+ 73,
+ 177,
+ 83
+ ],
+ "score": 0.86,
+ "content": "A \\land B )",
+ "type": "inline_equation",
+ "cross_page": true
+ },
+ {
+ "bbox": [
+ 177,
+ 72,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "). The latter operation is computationally simpler than checking if a file contains all",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 447,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 447,
+ 97
+ ],
+ "score": 1.0,
+ "content": "packages, since Codex can delete a file whenever a single package in the set appears.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 53.5,
+ "bbox_fs": [
+ 105,
+ 698,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 73,
+ 504,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 72,
+ 147,
+ 85
+ ],
+ "score": 1.0,
+ "content": "instead of",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 148,
+ 73,
+ 177,
+ 83
+ ],
+ "score": 0.86,
+ "content": "A \\land B )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 177,
+ 72,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "). The latter operation is computationally simpler than checking if a file contains all",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 105,
+ 83,
+ 447,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 83,
+ 447,
+ 97
+ ],
+ "score": 1.0,
+ "content": "packages, since Codex can delete a file whenever a single package in the set appears.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 100,
+ 505,
+ 166
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 146,
+ 112
+ ],
+ "score": 1.0,
+ "content": "In Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 99,
+ 157,
+ 113
+ ],
+ "score": 0.82,
+ "content": "\\textcircled { 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 100,
+ 505,
+ 112
+ ],
+ "score": 1.0,
+ "content": "we illustrate the breakdown of the errors Codex makes as a function of the number",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 111,
+ 506,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 473,
+ 123
+ ],
+ "score": 1.0,
+ "content": "of package imports in the prompt. We find that Codex erroneously deletes files on at least",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 111,
+ 493,
+ 122
+ ],
+ "score": 0.87,
+ "content": "80 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 111,
+ 506,
+ 123
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 121,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 121,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": "prompts when the number of package imports is at least three, despite producing a correct output",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 119,
+ 145
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 119,
+ 133,
+ 140,
+ 143
+ ],
+ "score": 0.86,
+ "content": "90 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "of prompts when the number of packages is at most two. Moreover, we find that Codex",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "increasingly errs by using only the first package as the problem gets more challenging (i.e. the number",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 333,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 333,
+ 167
+ ],
+ "score": 1.0,
+ "content": "of packages increases), as attribute substitution predicts.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 171,
+ 505,
+ 226
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 170,
+ 507,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 170,
+ 507,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Control experiments. To very that our findings generalize to different classes of realistic prompts,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 181,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 181,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "we test Codex on prompts containing a descriptive docstring beneath the function signature",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "delete_all_with_libraries(directory). We observe qualitatively similar results, though",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 203,
+ 504,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 203,
+ 358,
+ 216
+ ],
+ "score": 1.0,
+ "content": "we find more instances of low-impact errors; see Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 203,
+ 370,
+ 216
+ ],
+ "score": 0.53,
+ "content": "\\mathbf { \\bar { C } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 203,
+ 504,
+ 216
+ ],
+ "score": 1.0,
+ "content": "for details. Overall, our results",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 215,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 215,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "demonstrate how our framework can preemptively elicit high-impact errors, like erroneous deletions.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 241,
+ 180,
+ 254
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 240,
+ 181,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 240,
+ 181,
+ 257
+ ],
+ "score": 1.0,
+ "content": "6 Discussion",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 266,
+ 505,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "score": 1.0,
+ "content": "In this work, we identify and test for classes of errors that open-ended generation systems can make,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "using cognitive biases as motivation. To do so, we generate hypotheses for potential qualitative failure",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 301
+ ],
+ "score": 1.0,
+ "content": "modes, then construct transformations over prompts that elicit these failures. Our experiments uncover",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "deficiencies of Codex, CodeGen, and GPT-3, and elicit high-impact errors that are challenging to",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "undo. While we focus on a few specific failure modes, future work could apply our framework to",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "uncover additional failures. Moreover, our framework queries systems as a black-box, so it could be",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 388,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 388,
+ 344
+ ],
+ "score": 1.0,
+ "content": "used to quickly probe for errors in future systems as they are released.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 348,
+ 505,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "Some of our results highlight how optimizing likelihood could be at odds with human intent. For",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "example, over GitHub, programs may more often match their function signature than docstring",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 507,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 140,
+ 383
+ ],
+ "score": 1.0,
+ "content": "(Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 140,
+ 369,
+ 165,
+ 383
+ ],
+ "score": 0.39,
+ "content": "3 . 3 . 3 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 369,
+ 507,
+ 383
+ ],
+ "score": 1.0,
+ "content": ", or tend to complete to pass if the preceding function does (Section 3.3.1). Neverthe-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "less, our results elicit qualitative errors regardless of the “correct” behavior (i.e. even when what is",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 491,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 491,
+ 404
+ ],
+ "score": 1.0,
+ "content": "incorrect and correct flips), and demonstrate the importance of documenting qualitative failures.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 518
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "The reliability challenges posed by the open-ended generation systems that we study sometimes",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "also apply to classifiers. Some classification errors can be more costly than others [Oakden-Rayner",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 402,
+ 443
+ ],
+ "score": 1.0,
+ "content": "et al., 2020], classifiers may use irrelevant information to make predictions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 429,
+ 484,
+ 442
+ ],
+ "score": 0.62,
+ "content": "\\lVert \\overline { { \\mathrm { S a g a w a ~ e t ~ a l . } \\rVert \\mathbb { 2 0 } 2 0 } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 485,
+ 429,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 440,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 352,
+ 454
+ ],
+ "score": 1.0,
+ "content": "input-level transformations like universal adversarial triggers",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 441,
+ 438,
+ 453
+ ],
+ "score": 0.34,
+ "content": "\\rVert \\overline { { \\mathrm { W a l l a c e ~ e t ~ a l . } } } \\rVert \\overline { { 2 0 1 9 } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 439,
+ 440,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "and distribution",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 504,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 241,
+ 465
+ ],
+ "score": 1.0,
+ "content": "shifts [Hendrycks and Dietterich,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 451,
+ 267,
+ 464
+ ],
+ "score": 0.45,
+ "content": "\\boxed { 2 0 1 9 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 451,
+ 504,
+ 465
+ ],
+ "score": 1.0,
+ "content": "induce errors. However, while classification errors may be",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "succinctly summarized with a confusion matrix, generation errors cannot, since each output appears",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "infrequently. To tame the large output space, our transformations must induce categories of errors that",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "we can reliably measure. Despite this additional constraint, we are able to construct model-agnostic",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "transformations: we do not use the training data, model parameters, or even output logits. Our success",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 461,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 461,
+ 519
+ ],
+ "score": 1.0,
+ "content": "in this restricted setting demonstrates the comparative brittleness of completion systems.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 523,
+ 505,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "We present a method to systematically elicit errors from large language models. While we believe",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "our work is important to understand model behavior, bad actors could exploit the errors we reveal",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 558
+ ],
+ "score": 1.0,
+ "content": "(e.g. by deleting files on systems with a Codex back-end). Nevertheless, we introduce new robustness",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 556,
+ 503,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 503,
+ 568
+ ],
+ "score": 1.0,
+ "content": "challenges for developers and identify misuses of these models, which we feel supersedes this risk.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 571,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "As a subroutine in our experimental pipeline, we use cognitive biases as inspiration to identify",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "potential failure modes. This is an example of using a reference system—a system that is analogous to",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "the ML models we study in some meaningful way—to generate insights into ML systems [Steinhardt,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 131,
+ 617
+ ],
+ "score": 0.44,
+ "content": "\\boxed { 2 0 2 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 605,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "We use humans as the reference, focusing specifically on their susceptibility to cognitive",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "biases. Other references, such as complex systems or evolution, may uncover new errors and insights.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 639
+ ],
+ "score": 1.0,
+ "content": "Moreover, ML systems could additionally err in ways that known systems do not, so it will also be",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "useful to have intrinsic methods for characterizing model errors. Overall, our work underscores the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 488,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 488,
+ 660
+ ],
+ "score": 1.0,
+ "content": "need for more extensive testing of generative ML systems before their widespread deployment.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 43.5
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 675,
+ 207,
+ 688
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 208,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 208,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Acknowledgements",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "We thank the anonymous reviewers, Ruiqi Zhong, Jean-Stanislas Denain, Aditi Raghunathan, Jessy",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Lin, and Lawrence Chan for feedback. This work was supported by NSF Award Grant no. 1804794.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49.5
+ }
+ ],
+ "page_idx": 9,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 301,
+ 742,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 14
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 73,
+ 504,
+ 95
+ ],
+ "lines": [],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 72,
+ 505,
+ 97
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 100,
+ 505,
+ 166
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 99,
+ 505,
+ 113
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 146,
+ 112
+ ],
+ "score": 1.0,
+ "content": "In Figure",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 146,
+ 99,
+ 157,
+ 113
+ ],
+ "score": 0.82,
+ "content": "\\textcircled { 6 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 157,
+ 100,
+ 505,
+ 112
+ ],
+ "score": 1.0,
+ "content": "we illustrate the breakdown of the errors Codex makes as a function of the number",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 105,
+ 111,
+ 506,
+ 123
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 111,
+ 473,
+ 123
+ ],
+ "score": 1.0,
+ "content": "of package imports in the prompt. We find that Codex erroneously deletes files on at least",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 473,
+ 111,
+ 493,
+ 122
+ ],
+ "score": 0.87,
+ "content": "80 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 493,
+ 111,
+ 506,
+ 123
+ ],
+ "score": 1.0,
+ "content": "of",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 104,
+ 121,
+ 506,
+ 135
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 121,
+ 506,
+ 135
+ ],
+ "score": 1.0,
+ "content": "prompts when the number of package imports is at least three, despite producing a correct output",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 105,
+ 132,
+ 506,
+ 145
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 132,
+ 119,
+ 145
+ ],
+ "score": 1.0,
+ "content": "on",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 119,
+ 133,
+ 140,
+ 143
+ ],
+ "score": 0.86,
+ "content": "90 \\%",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 140,
+ 132,
+ 506,
+ 145
+ ],
+ "score": 1.0,
+ "content": "of prompts when the number of packages is at most two. Moreover, we find that Codex",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 144,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "increasingly errs by using only the first package as the problem gets more challenging (i.e. the number",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 105,
+ 155,
+ 333,
+ 167
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 155,
+ 333,
+ 167
+ ],
+ "score": 1.0,
+ "content": "of packages increases), as attribute substitution predicts.",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ }
+ ],
+ "index": 4.5,
+ "bbox_fs": [
+ 104,
+ 99,
+ 506,
+ 167
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 171,
+ 505,
+ 226
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 170,
+ 507,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 170,
+ 507,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Control experiments. To very that our findings generalize to different classes of realistic prompts,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 106,
+ 181,
+ 505,
+ 194
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 181,
+ 505,
+ 194
+ ],
+ "score": 1.0,
+ "content": "we test Codex on prompts containing a descriptive docstring beneath the function signature",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 192,
+ 505,
+ 205
+ ],
+ "score": 1.0,
+ "content": "delete_all_with_libraries(directory). We observe qualitatively similar results, though",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 106,
+ 203,
+ 504,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 203,
+ 358,
+ 216
+ ],
+ "score": 1.0,
+ "content": "we find more instances of low-impact errors; see Appendix",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 358,
+ 203,
+ 370,
+ 216
+ ],
+ "score": 0.53,
+ "content": "\\mathbf { \\bar { C } }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 370,
+ 203,
+ 504,
+ 216
+ ],
+ "score": 1.0,
+ "content": "for details. Overall, our results",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 106,
+ 215,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 215,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "demonstrate how our framework can preemptively elicit high-impact errors, like erroneous deletions.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 105,
+ 170,
+ 507,
+ 227
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 241,
+ 180,
+ 254
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 240,
+ 181,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 240,
+ 181,
+ 257
+ ],
+ "score": 1.0,
+ "content": "6 Discussion",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 266,
+ 505,
+ 343
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 266,
+ 506,
+ 279
+ ],
+ "score": 1.0,
+ "content": "In this work, we identify and test for classes of errors that open-ended generation systems can make,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 276,
+ 505,
+ 290
+ ],
+ "score": 1.0,
+ "content": "using cognitive biases as motivation. To do so, we generate hypotheses for potential qualitative failure",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 301
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 288,
+ 506,
+ 301
+ ],
+ "score": 1.0,
+ "content": "modes, then construct transformations over prompts that elicit these failures. Our experiments uncover",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 298,
+ 505,
+ 312
+ ],
+ "score": 1.0,
+ "content": "deficiencies of Codex, CodeGen, and GPT-3, and elicit high-impact errors that are challenging to",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 309,
+ 505,
+ 322
+ ],
+ "score": 1.0,
+ "content": "undo. While we focus on a few specific failure modes, future work could apply our framework to",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 333
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 321,
+ 505,
+ 333
+ ],
+ "score": 1.0,
+ "content": "uncover additional failures. Moreover, our framework queries systems as a black-box, so it could be",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 105,
+ 332,
+ 388,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 332,
+ 388,
+ 344
+ ],
+ "score": 1.0,
+ "content": "used to quickly probe for errors in future systems as they are released.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 105,
+ 266,
+ 506,
+ 344
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 348,
+ 505,
+ 403
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 348,
+ 505,
+ 360
+ ],
+ "score": 1.0,
+ "content": "Some of our results highlight how optimizing likelihood could be at odds with human intent. For",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 372
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 358,
+ 506,
+ 372
+ ],
+ "score": 1.0,
+ "content": "example, over GitHub, programs may more often match their function signature than docstring",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 105,
+ 369,
+ 507,
+ 383
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 369,
+ 140,
+ 383
+ ],
+ "score": 1.0,
+ "content": "(Section",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 140,
+ 369,
+ 165,
+ 383
+ ],
+ "score": 0.39,
+ "content": "3 . 3 . 3 )",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 165,
+ 369,
+ 507,
+ 383
+ ],
+ "score": 1.0,
+ "content": ", or tend to complete to pass if the preceding function does (Section 3.3.1). Neverthe-",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 381,
+ 505,
+ 393
+ ],
+ "score": 1.0,
+ "content": "less, our results elicit qualitative errors regardless of the “correct” behavior (i.e. even when what is",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 106,
+ 392,
+ 491,
+ 404
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 392,
+ 491,
+ 404
+ ],
+ "score": 1.0,
+ "content": "incorrect and correct flips), and demonstrate the importance of documenting qualitative failures.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 105,
+ 348,
+ 507,
+ 404
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 408,
+ 505,
+ 518
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 407,
+ 505,
+ 420
+ ],
+ "score": 1.0,
+ "content": "The reliability challenges posed by the open-ended generation systems that we study sometimes",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 432
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 419,
+ 506,
+ 432
+ ],
+ "score": 1.0,
+ "content": "also apply to classifiers. Some classification errors can be more costly than others [Oakden-Rayner",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 443
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 402,
+ 443
+ ],
+ "score": 1.0,
+ "content": "et al., 2020], classifiers may use irrelevant information to make predictions",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 402,
+ 429,
+ 484,
+ 442
+ ],
+ "score": 0.62,
+ "content": "\\lVert \\overline { { \\mathrm { S a g a w a ~ e t ~ a l . } \\rVert \\mathbb { 2 0 } 2 0 } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 485,
+ 429,
+ 505,
+ 443
+ ],
+ "score": 1.0,
+ "content": ", and",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 105,
+ 440,
+ 506,
+ 454
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 440,
+ 352,
+ 454
+ ],
+ "score": 1.0,
+ "content": "input-level transformations like universal adversarial triggers",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 353,
+ 441,
+ 438,
+ 453
+ ],
+ "score": 0.34,
+ "content": "\\rVert \\overline { { \\mathrm { W a l l a c e ~ e t ~ a l . } } } \\rVert \\overline { { 2 0 1 9 } } \\rVert",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 439,
+ 440,
+ 506,
+ 454
+ ],
+ "score": 1.0,
+ "content": "and distribution",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 105,
+ 451,
+ 504,
+ 465
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 451,
+ 241,
+ 465
+ ],
+ "score": 1.0,
+ "content": "shifts [Hendrycks and Dietterich,",
+ "type": "text"
+ },
+ {
+ "bbox": [
+ 241,
+ 451,
+ 267,
+ 464
+ ],
+ "score": 0.45,
+ "content": "\\boxed { 2 0 1 9 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 267,
+ 451,
+ 504,
+ 465
+ ],
+ "score": 1.0,
+ "content": "induce errors. However, while classification errors may be",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "succinctly summarized with a confusion matrix, generation errors cannot, since each output appears",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 473,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "infrequently. To tame the large output space, our transformations must induce categories of errors that",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 484,
+ 505,
+ 497
+ ],
+ "score": 1.0,
+ "content": "we can reliably measure. Despite this additional constraint, we are able to construct model-agnostic",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 507
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 496,
+ 505,
+ 507
+ ],
+ "score": 1.0,
+ "content": "transformations: we do not use the training data, model parameters, or even output logits. Our success",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 105,
+ 505,
+ 461,
+ 519
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 505,
+ 461,
+ 519
+ ],
+ "score": 1.0,
+ "content": "in this restricted setting demonstrates the comparative brittleness of completion systems.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 105,
+ 407,
+ 506,
+ 519
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 523,
+ 505,
+ 567
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 523,
+ 505,
+ 536
+ ],
+ "score": 1.0,
+ "content": "We present a method to systematically elicit errors from large language models. While we believe",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "our work is important to understand model behavior, bad actors could exploit the errors we reveal",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 558
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 545,
+ 506,
+ 558
+ ],
+ "score": 1.0,
+ "content": "(e.g. by deleting files on systems with a Codex back-end). Nevertheless, we introduce new robustness",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 106,
+ 556,
+ 503,
+ 568
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 556,
+ 503,
+ 568
+ ],
+ "score": 1.0,
+ "content": "challenges for developers and identify misuses of these models, which we feel supersedes this risk.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 523,
+ 506,
+ 568
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 571,
+ 505,
+ 660
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 571,
+ 506,
+ 585
+ ],
+ "score": 1.0,
+ "content": "As a subroutine in our experimental pipeline, we use cognitive biases as inspiration to identify",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 595
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 583,
+ 506,
+ 595
+ ],
+ "score": 1.0,
+ "content": "potential failure modes. This is an example of using a reference system—a system that is analogous to",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 606
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 593,
+ 505,
+ 606
+ ],
+ "score": 1.0,
+ "content": "the ML models we study in some meaningful way—to generate insights into ML systems [Steinhardt,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 105,
+ 604,
+ 506,
+ 617
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 604,
+ 131,
+ 617
+ ],
+ "score": 0.44,
+ "content": "\\boxed { 2 0 2 2 }",
+ "type": "inline_equation"
+ },
+ {
+ "bbox": [
+ 135,
+ 605,
+ 506,
+ 617
+ ],
+ "score": 1.0,
+ "content": "We use humans as the reference, focusing specifically on their susceptibility to cognitive",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 615,
+ 506,
+ 628
+ ],
+ "score": 1.0,
+ "content": "biases. Other references, such as complex systems or evolution, may uncover new errors and insights.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 639
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 626,
+ 506,
+ 639
+ ],
+ "score": 1.0,
+ "content": "Moreover, ML systems could additionally err in ways that known systems do not, so it will also be",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 649
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 638,
+ 505,
+ 649
+ ],
+ "score": 1.0,
+ "content": "useful to have intrinsic methods for characterizing model errors. Overall, our work underscores the",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 105,
+ 648,
+ 488,
+ 660
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 648,
+ 488,
+ 660
+ ],
+ "score": 1.0,
+ "content": "need for more extensive testing of generative ML systems before their widespread deployment.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 105,
+ 571,
+ 506,
+ 660
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 108,
+ 675,
+ 207,
+ 688
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 208,
+ 691
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 673,
+ 208,
+ 691
+ ],
+ "score": 1.0,
+ "content": "Acknowledgements",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ }
+ ],
+ "index": 48
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 698,
+ 505,
+ 713
+ ],
+ "score": 1.0,
+ "content": "We thank the anonymous reviewers, Ruiqi Zhong, Jean-Stanislas Denain, Aditi Raghunathan, Jessy",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 710,
+ 506,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Lin, and Lawrence Chan for feedback. This work was supported by NSF Award Grant no. 1804794.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ }
+ ],
+ "index": 49.5,
+ "bbox_fs": [
+ 105,
+ 698,
+ 506,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 71,
+ 163,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 90,
+ 505,
+ 123
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 103
+ ],
+ "score": 1.0,
+ "content": "Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchel. On the dan-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 101,
+ 507,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 101,
+ 507,
+ 115
+ ],
+ "score": 1.0,
+ "content": "gers of stochastic parrots: Can language models be too big? In ACM Conference on Fairness,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 112,
+ 315,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 112,
+ 315,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Accountability, and Transparency (FAccT), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 130,
+ 504,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. Stereotyping",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 142,
+ 505,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 142,
+ 505,
+ 154
+ ],
+ "score": 1.0,
+ "content": "norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Association for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 117,
+ 153,
+ 279,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 153,
+ 279,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Computational Linguistics (ACL), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 167,
+ 506,
+ 389
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Ste-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 214,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 214,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "fano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 116,
+ 226,
+ 505,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 226,
+ 505,
+ 237
+ ],
+ "score": 1.0,
+ "content": "Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 236,
+ 506,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 236,
+ 506,
+ 250
+ ],
+ "score": 1.0,
+ "content": "Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 114,
+ 246,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 246,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 256,
+ 507,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 256,
+ 507,
+ 272
+ ],
+ "score": 1.0,
+ "content": "Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 269,
+ 506,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 269,
+ 506,
+ 281
+ ],
+ "score": 1.0,
+ "content": "Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 279,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 279,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 303
+ ],
+ "score": 1.0,
+ "content": "Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Car-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 301,
+ 507,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 301,
+ 507,
+ 315
+ ],
+ "score": 1.0,
+ "content": "los Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 312,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 312,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 324,
+ 506,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 324,
+ 506,
+ 336
+ ],
+ "score": 1.0,
+ "content": "Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Ré, Dorsa",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 334,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 334,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 344,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 344,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 114,
+ 354,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 354,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei Zaharia, Michael Zhang,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 114,
+ 366,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 366,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 378,
+ 464,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 378,
+ 464,
+ 390
+ ],
+ "score": 1.0,
+ "content": "opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 396,
+ 506,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 506,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 506,
+ 408
+ ],
+ "score": 1.0,
+ "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhari-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 407,
+ 506,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 407,
+ 506,
+ 420
+ ],
+ "score": 1.0,
+ "content": "wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 417,
+ 507,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 417,
+ 507,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 114,
+ 428,
+ 507,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 428,
+ 507,
+ 442
+ ],
+ "score": 1.0,
+ "content": "Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 439,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 439,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 462,
+ 219,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 462,
+ 219,
+ 473
+ ],
+ "score": 1.0,
+ "content": "arXiv:2005.14165, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 480,
+ 503,
+ 514
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 479,
+ 504,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 479,
+ 504,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 492,
+ 504,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 492,
+ 504,
+ 504
+ ],
+ "score": 1.0,
+ "content": "Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Colin Raffel.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 117,
+ 502,
+ 492,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 502,
+ 492,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Extracting training data from large language models. arXiv preprint arXiv:2012.07805, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 521,
+ 506,
+ 641
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 532,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 532,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri,",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 541,
+ 507,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 541,
+ 507,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 552,
+ 507,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 552,
+ 507,
+ 567
+ ],
+ "score": 1.0,
+ "content": "Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 575,
+ 507,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 575,
+ 507,
+ 588
+ ],
+ "score": 1.0,
+ "content": "Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 586,
+ 507,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 586,
+ 507,
+ 599
+ ],
+ "score": 1.0,
+ "content": "Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 596,
+ 507,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 596,
+ 507,
+ 610
+ ],
+ "score": 1.0,
+ "content": "Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 608,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 608,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 116,
+ 630,
+ 439,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 630,
+ 439,
+ 642
+ ],
+ "score": 1.0,
+ "content": "large language models trained on code. arXiv preprint arXiv:2107.03374, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 42
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 648,
+ 506,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 507,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 507,
+ 661
+ ],
+ "score": 1.0,
+ "content": "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser,",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 115,
+ 660,
+ 505,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 660,
+ 505,
+ 671
+ ],
+ "score": 1.0,
+ "content": "Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 114,
+ 670,
+ 507,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 670,
+ 507,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 114,
+ 681,
+ 143,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 681,
+ 143,
+ 693
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 699,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. ELI5:",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 115,
+ 710,
+ 481,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 710,
+ 481,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Long form question answering. In Association for Computational Linguistics (ACL), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 52.5
+ }
+ ],
+ "page_idx": 10,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 310,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "score": 1.0,
+ "content": "",
+ "type": "text",
+ "height": 14,
+ "width": 13
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 71,
+ 163,
+ 84
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 70,
+ 165,
+ 86
+ ],
+ "score": 1.0,
+ "content": "References",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ }
+ ],
+ "index": 0
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 90,
+ 505,
+ 123
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 103
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 89,
+ 506,
+ 103
+ ],
+ "score": 1.0,
+ "content": "Emily Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchel. On the dan-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 101,
+ 507,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 101,
+ 507,
+ 115
+ ],
+ "score": 1.0,
+ "content": "gers of stochastic parrots: Can language models be too big? In ACM Conference on Fairness,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 112,
+ 315,
+ 124
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 112,
+ 315,
+ 124
+ ],
+ "score": 1.0,
+ "content": "Accountability, and Transparency (FAccT), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2,
+ "bbox_fs": [
+ 106,
+ 89,
+ 507,
+ 124
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 130,
+ 504,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 128,
+ 506,
+ 144
+ ],
+ "score": 1.0,
+ "content": "Su Lin Blodgett, Gilsinia Lopez, Alexandra Olteanu, Robert Sim, and Hanna Wallach. Stereotyping",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 142,
+ 505,
+ 154
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 142,
+ 505,
+ 154
+ ],
+ "score": 1.0,
+ "content": "norwegian salmon: An inventory of pitfalls in fairness benchmark datasets. In Association for",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 117,
+ 153,
+ 279,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 153,
+ 279,
+ 165
+ ],
+ "score": 1.0,
+ "content": "Computational Linguistics (ACL), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 105,
+ 128,
+ 506,
+ 165
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 167,
+ 506,
+ 389
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 169,
+ 506,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Rishi Bommasani, Drew A. Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx,",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "Michael S. Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, Erik Brynjolfsson,",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 205
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 205
+ ],
+ "score": 1.0,
+ "content": "Shyamal Buch, Dallas Card, Rodrigo Castellon, Niladri Chatterji, Annie Chen, Kathleen Creel,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Jared Quincy Davis, Dorottya Demszky, Chris Donahue, Moussa Doumbouya, Esin Durmus, Ste-",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 214,
+ 506,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 214,
+ 506,
+ 227
+ ],
+ "score": 1.0,
+ "content": "fano Ermon, John Etchemendy, Kawin Ethayarajh, Li Fei-Fei, Chelsea Finn, Trevor Gale, Lauren",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 116,
+ 226,
+ 505,
+ 237
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 226,
+ 505,
+ 237
+ ],
+ "score": 1.0,
+ "content": "Gillespie, Karan Goel, Noah Goodman, Shelby Grossman, Neel Guha, Tatsunori Hashimoto, Peter",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 236,
+ 506,
+ 250
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 236,
+ 506,
+ 250
+ ],
+ "score": 1.0,
+ "content": "Henderson, John Hewitt, Daniel E. Ho, Jenny Hong, Kyle Hsu, Jing Huang, Thomas Icard, Saahil",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 114,
+ 246,
+ 506,
+ 261
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 246,
+ 506,
+ 261
+ ],
+ "score": 1.0,
+ "content": "Jain, Dan Jurafsky, Pratyusha Kalluri, Siddharth Karamcheti, Geoff Keeling, Fereshte Khani,",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ },
+ {
+ "bbox": [
+ 115,
+ 256,
+ 507,
+ 272
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 256,
+ 507,
+ 272
+ ],
+ "score": 1.0,
+ "content": "Omar Khattab, Pang Wei Koh, Mark Krass, Ranjay Krishna, Rohith Kuditipudi, Ananya Kumar,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 269,
+ 506,
+ 281
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 269,
+ 506,
+ 281
+ ],
+ "score": 1.0,
+ "content": "Faisal Ladhak, Mina Lee, Tony Lee, Jure Leskovec, Isabelle Levent, Xiang Lisa Li, Xuechen",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 279,
+ 506,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 279,
+ 506,
+ 293
+ ],
+ "score": 1.0,
+ "content": "Li, Tengyu Ma, Ali Malik, Christopher D. Manning, Suvir Mirchandani, Eric Mitchell, Zanele",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 303
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 303
+ ],
+ "score": 1.0,
+ "content": "Munyikwa, Suraj Nair, Avanika Narayan, Deepak Narayanan, Ben Newman, Allen Nie, Juan Car-",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 301,
+ 507,
+ 315
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 301,
+ 507,
+ 315
+ ],
+ "score": 1.0,
+ "content": "los Niebles, Hamed Nilforoshan, Julian Nyarko, Giray Ogut, Laurel Orr, Isabel Papadimitriou,",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 312,
+ 506,
+ 325
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 312,
+ 506,
+ 325
+ ],
+ "score": 1.0,
+ "content": "Joon Sung Park, Chris Piech, Eva Portelance, Christopher Potts, Aditi Raghunathan, Rob Reich,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 324,
+ 506,
+ 336
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 324,
+ 506,
+ 336
+ ],
+ "score": 1.0,
+ "content": "Hongyu Ren, Frieda Rong, Yusuf Roohani, Camilo Ruiz, Jack Ryan, Christopher Ré, Dorsa",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 334,
+ 505,
+ 347
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 334,
+ 505,
+ 347
+ ],
+ "score": 1.0,
+ "content": "Sadigh, Shiori Sagawa, Keshav Santhanam, Andy Shih, Krishnan Srinivasan, Alex Tamkin, Rohan",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 115,
+ 344,
+ 506,
+ 358
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 344,
+ 506,
+ 358
+ ],
+ "score": 1.0,
+ "content": "Taori, Armin W. Thomas, Florian Tramèr, Rose E. Wang, William Wang, Bohan Wu, Jiajun Wu,",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 114,
+ 354,
+ 506,
+ 370
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 354,
+ 506,
+ 370
+ ],
+ "score": 1.0,
+ "content": "Yuhuai Wu, Sang Michael Xie, Michihiro Yasunaga, Jiaxuan You, Matei Zaharia, Michael Zhang,",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 114,
+ 366,
+ 506,
+ 380
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 366,
+ 506,
+ 380
+ ],
+ "score": 1.0,
+ "content": "Tianyi Zhang, Xikun Zhang, Yuhui Zhang, Lucia Zheng, Kaitlyn Zhou, and Percy Liang. On the",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 378,
+ 464,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 378,
+ 464,
+ 390
+ ],
+ "score": 1.0,
+ "content": "opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 105,
+ 169,
+ 507,
+ 390
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 396,
+ 506,
+ 473
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 506,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 396,
+ 506,
+ 408
+ ],
+ "score": 1.0,
+ "content": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhari-",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 116,
+ 407,
+ 506,
+ 420
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 407,
+ 506,
+ 420
+ ],
+ "score": 1.0,
+ "content": "wal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal,",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 417,
+ 507,
+ 431
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 417,
+ 507,
+ 431
+ ],
+ "score": 1.0,
+ "content": "Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ },
+ {
+ "bbox": [
+ 114,
+ 428,
+ 507,
+ 442
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 428,
+ 507,
+ 442
+ ],
+ "score": 1.0,
+ "content": "Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin,",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 115,
+ 439,
+ 506,
+ 453
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 439,
+ 506,
+ 453
+ ],
+ "score": 1.0,
+ "content": "Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 451,
+ 506,
+ 464
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 451,
+ 506,
+ 464
+ ],
+ "score": 1.0,
+ "content": "Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 462,
+ 219,
+ 473
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 462,
+ 219,
+ 473
+ ],
+ "score": 1.0,
+ "content": "arXiv:2005.14165, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 106,
+ 396,
+ 507,
+ 473
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 480,
+ 503,
+ 514
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 479,
+ 504,
+ 493
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 479,
+ 504,
+ 493
+ ],
+ "score": 1.0,
+ "content": "Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert-Voss, Katherine",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 492,
+ 504,
+ 504
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 492,
+ 504,
+ 504
+ ],
+ "score": 1.0,
+ "content": "Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, Alina Oprea, and Colin Raffel.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 117,
+ 502,
+ 492,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 502,
+ 492,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Extracting training data from large language models. arXiv preprint arXiv:2012.07805, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 107,
+ 479,
+ 504,
+ 515
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 521,
+ 506,
+ 641
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 534
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 520,
+ 506,
+ 534
+ ],
+ "score": 1.0,
+ "content": "Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 116,
+ 532,
+ 506,
+ 544
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 532,
+ 506,
+ 544
+ ],
+ "score": 1.0,
+ "content": "Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, Alex Ray, Raul Puri,",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 115,
+ 541,
+ 507,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 541,
+ 507,
+ 556
+ ],
+ "score": 1.0,
+ "content": "Gretchen Krueger, Michael Petrov, Heidy Khlaaf, Girish Sastry, Pamela Mishkin, Brooke Chan,",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 115,
+ 552,
+ 507,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 552,
+ 507,
+ 567
+ ],
+ "score": 1.0,
+ "content": "Scott Gray, Nick Ryder, Mikhail Pavlov, Alethea Power, Lukasz Kaiser, Mohammad Bavarian,",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 564,
+ 506,
+ 577
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 564,
+ 506,
+ 577
+ ],
+ "score": 1.0,
+ "content": "Clemens Winter, Philippe Tillet, Felipe Petroski Such, Dave Cummings, Matthias Plappert, Fotios",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 575,
+ 507,
+ 588
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 575,
+ 507,
+ 588
+ ],
+ "score": 1.0,
+ "content": "Chantzis, Elizabeth Barnes, Ariel Herbert-Voss, William Hebgen Guss, Alex Nichol, Alex Paino,",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 586,
+ 507,
+ 599
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 586,
+ 507,
+ 599
+ ],
+ "score": 1.0,
+ "content": "Nikolas Tezak, Jie Tang, Igor Babuschkin, Suchir Balaji, Shantanu Jain, William Saunders,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 596,
+ 507,
+ 610
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 596,
+ 507,
+ 610
+ ],
+ "score": 1.0,
+ "content": "Christopher Hesse, Andrew N. Carr, Jan Leike, Josh Achiam, Vedant Misra, Evan Morikawa,",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 608,
+ 505,
+ 621
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 608,
+ 505,
+ 621
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Matthew Knight, Miles Brundage, Mira Murati, Katie Mayer, Peter Welinder, Bob",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 633
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 618,
+ 506,
+ 633
+ ],
+ "score": 1.0,
+ "content": "McGrew, Dario Amodei, Sam McCandlish, Ilya Sutskever, and Wojciech Zaremba. Evaluating",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 116,
+ 630,
+ 439,
+ 642
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 630,
+ 439,
+ 642
+ ],
+ "score": 1.0,
+ "content": "large language models trained on code. arXiv preprint arXiv:2107.03374, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 42,
+ "bbox_fs": [
+ 105,
+ 520,
+ 507,
+ 642
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 648,
+ 506,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 507,
+ 661
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 647,
+ 507,
+ 661
+ ],
+ "score": 1.0,
+ "content": "Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser,",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 115,
+ 660,
+ 505,
+ 671
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 660,
+ 505,
+ 671
+ ],
+ "score": 1.0,
+ "content": "Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ },
+ {
+ "bbox": [
+ 114,
+ 670,
+ 507,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 670,
+ 507,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Schulman. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168,",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 114,
+ 681,
+ 143,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 681,
+ 143,
+ 693
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 49.5,
+ "bbox_fs": [
+ 105,
+ 647,
+ 507,
+ 693
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 699,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 108,
+ 700,
+ 505,
+ 712
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 108,
+ 700,
+ 505,
+ 712
+ ],
+ "score": 1.0,
+ "content": "Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jason Weston, and Michael Auli. ELI5:",
+ "type": "text"
+ }
+ ],
+ "index": 52
+ },
+ {
+ "bbox": [
+ 115,
+ 710,
+ 481,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 710,
+ 481,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Long form question answering. In Association for Computational Linguistics (ACL), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 53
+ }
+ ],
+ "index": 52.5,
+ "bbox_fs": [
+ 108,
+ 700,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 504,
+ 106
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "Saadia Gabriel, Asli Celikyilmaz, Rahul Jha, Yejin Choi, and Jianfeng Gao. GO FIGURE: A meta",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "evaluation of factuality in summarization. In Findings of the Association for Computational",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 95,
+ 264,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 95,
+ 264,
+ 107
+ ],
+ "score": 1.0,
+ "content": "Linguistics (Findings of ACL), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Real-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 125,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 125,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "toxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 137,
+ 218,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 137,
+ 218,
+ 147
+ ],
+ "score": 1.0,
+ "content": "arXiv:2009.11462, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 156,
+ 504,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Sophie Groenwold, Lily Ou, Aesha Parekh, Samhita Honnavalli, Sharon Levy, Diba Mirza, and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 168,
+ 506,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 168,
+ 506,
+ 180
+ ],
+ "score": 1.0,
+ "content": "William Yang Wang. Investigating african-american vernacular english in transformer-based text",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 179,
+ 456,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 179,
+ 456,
+ 191
+ ],
+ "score": 1.0,
+ "content": "generation. In Empirical Methods in Natural Language Processing (EMNLP), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 198,
+ 504,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 209,
+ 507,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 209,
+ 507,
+ 223
+ ],
+ "score": 1.0,
+ "content": "corruptions and perturbations. In International Conference on Learning Representations (ICLR),",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 219,
+ 144,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 219,
+ 144,
+ 232
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 240,
+ 504,
+ 275
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 253
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks, Steven Basart, Saurav Kadavath, Mantas Mazeika, Akul Arora, Ethan Guo, Collin",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 250,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 250,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "Burns, Samir Puranik, Horace He, Dawn Song, and Jacob Steinhardt. Measuring coding challenge",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 263,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 263,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "competence with APPS. In Advances in Neural Information Processing Systems (NeurIPS), 2021a.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 282,
+ 504,
+ 316
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 294,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 294,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "Steinhardt. Measuring massive multitask language understanding. In International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 116,
+ 305,
+ 284,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 305,
+ 284,
+ 317
+ ],
+ "score": 1.0,
+ "content": "Learning Representations (ICLR), 2021b.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 324,
+ 506,
+ 358
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. In",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 347,
+ 403,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 347,
+ 403,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Advances in Neural Information Processing Systems (NeurIPS), 2021c.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 366,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 365,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 365,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Abigail Z. Jacobs and Hanna Wallach. Measurement and fairness. In ACM Conference on Fairness,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 114,
+ 378,
+ 315,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 378,
+ 315,
+ 390
+ ],
+ "score": 1.0,
+ "content": "Accountability, and Transparency (FAccT), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 397,
+ 505,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Karen E. Jacowitz and Daniel Kahneman. Measures of anchoring in estimation tasks. Personality",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 409,
+ 350,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 409,
+ 350,
+ 421
+ ],
+ "score": 1.0,
+ "content": "and Social Psychology Bulletin, 21(11):1161–1166, 1995.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 428,
+ 505,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "Daniel Kahneman and Shane Frederick. Representativeness revisited: Attribute substitution in",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 440,
+ 507,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 440,
+ 507,
+ 452
+ ],
+ "score": 1.0,
+ "content": "intuitive judgment. In Heuristics and Biases: The Psychology of Intuitive Judgement, pages 49–81.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 114,
+ 450,
+ 143,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 450,
+ 143,
+ 463
+ ],
+ "score": 1.0,
+ "content": "2002.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 470,
+ 504,
+ 494
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "Kalpesh Krishna, Aurko Roy, and Mohit Iyyer. Hurdles to progress in long-form question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 481,
+ 434,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 481,
+ 434,
+ 496
+ ],
+ "score": 1.0,
+ "content": "In North American Association for Computational Linguistics (NAACL), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 501,
+ 503,
+ 525
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 116,
+ 513,
+ 327,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 513,
+ 327,
+ 525
+ ],
+ "score": 1.0,
+ "content": "falsehoods. arXiv preprint arXiv:2109.07958, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 544,
+ 459,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 544,
+ 459,
+ 556
+ ],
+ "score": 1.0,
+ "content": "makes good in-context examples for GPT-3. arXiv preprint arXiv:2101.06804, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 564,
+ 464,
+ 576
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 563,
+ 464,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 563,
+ 464,
+ 578
+ ],
+ "score": 1.0,
+ "content": "David E. Meyer. Semantic priming well established. Science, 345(6196):523–523, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 607
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. Stereoset: Measuring stereotypical bias in pretrained",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 595,
+ 352,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 595,
+ 352,
+ 608
+ ],
+ "score": 1.0,
+ "content": "language models. arXiv preprint arXiv:2004.09456, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huam Wang, Yingbo Zhou, Silvio Savarese,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "and Caiming Xiong. A conversational paradigm for program synthesis. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 116,
+ 637,
+ 219,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 637,
+ 219,
+ 650
+ ],
+ "score": 1.0,
+ "content": "arXiv:2203.13474, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 657,
+ 504,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 670
+ ],
+ "score": 1.0,
+ "content": "Luke Oakden-Rayner, Jared Dunnmon, Gustavo Carneiro, and Christopher Ré. Hidden stratification",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 668,
+ 505,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 668,
+ 505,
+ 682
+ ],
+ "score": 1.0,
+ "content": "causes clinically meaningful failures in machine learning for medical imaging. In Proceedings of",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 679,
+ 438,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 679,
+ 438,
+ 693
+ ],
+ "score": 1.0,
+ "content": "the ACM Conference on Health, Inference, and Learning, pages 151–159, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 502,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 116,
+ 711,
+ 398,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 711,
+ 398,
+ 723
+ ],
+ "score": 1.0,
+ "content": "models are unsupervised multitask learners. OpenAI Blog, 1(8), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5
+ }
+ ],
+ "page_idx": 11,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "12",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 72,
+ 504,
+ 106
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 73,
+ 505,
+ 85
+ ],
+ "score": 1.0,
+ "content": "Saadia Gabriel, Asli Celikyilmaz, Rahul Jha, Yejin Choi, and Jianfeng Gao. GO FIGURE: A meta",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 83,
+ 505,
+ 96
+ ],
+ "score": 1.0,
+ "content": "evaluation of factuality in summarization. In Findings of the Association for Computational",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 116,
+ 95,
+ 264,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 95,
+ 264,
+ 107
+ ],
+ "score": 1.0,
+ "content": "Linguistics (Findings of ACL), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ }
+ ],
+ "index": 1,
+ "bbox_fs": [
+ 106,
+ 73,
+ 505,
+ 107
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 148
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 127
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 114,
+ 506,
+ 127
+ ],
+ "score": 1.0,
+ "content": "Samuel Gehman, Suchin Gururangan, Maarten Sap, Yejin Choi, and Noah A Smith. Real-",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 116,
+ 125,
+ 505,
+ 137
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 125,
+ 505,
+ 137
+ ],
+ "score": 1.0,
+ "content": "toxicityprompts: Evaluating neural toxic degeneration in language models. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 137,
+ 218,
+ 147
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 137,
+ 218,
+ 147
+ ],
+ "score": 1.0,
+ "content": "arXiv:2009.11462, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 4,
+ "bbox_fs": [
+ 106,
+ 114,
+ 506,
+ 147
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 156,
+ 504,
+ 190
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 169
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 156,
+ 506,
+ 169
+ ],
+ "score": 1.0,
+ "content": "Sophie Groenwold, Lily Ou, Aesha Parekh, Samhita Honnavalli, Sharon Levy, Diba Mirza, and",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 116,
+ 168,
+ 506,
+ 180
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 168,
+ 506,
+ 180
+ ],
+ "score": 1.0,
+ "content": "William Yang Wang. Investigating african-american vernacular english in transformer-based text",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 179,
+ 456,
+ 191
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 179,
+ 456,
+ 191
+ ],
+ "score": 1.0,
+ "content": "generation. In Empirical Methods in Natural Language Processing (EMNLP), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 105,
+ 156,
+ 506,
+ 191
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 198,
+ 504,
+ 231
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 199,
+ 505,
+ 211
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 209,
+ 507,
+ 223
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 209,
+ 507,
+ 223
+ ],
+ "score": 1.0,
+ "content": "corruptions and perturbations. In International Conference on Learning Representations (ICLR),",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 115,
+ 219,
+ 144,
+ 232
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 219,
+ 144,
+ 232
+ ],
+ "score": 1.0,
+ "content": "2019.",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ }
+ ],
+ "index": 10,
+ "bbox_fs": [
+ 106,
+ 199,
+ 507,
+ 232
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 240,
+ 504,
+ 275
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 253
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 240,
+ 505,
+ 253
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks, Steven Basart, Saurav Kadavath, Mantas Mazeika, Akul Arora, Ethan Guo, Collin",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 115,
+ 250,
+ 506,
+ 265
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 250,
+ 506,
+ 265
+ ],
+ "score": 1.0,
+ "content": "Burns, Samir Puranik, Horace He, Dawn Song, and Jacob Steinhardt. Measuring coding challenge",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 263,
+ 506,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 263,
+ 506,
+ 275
+ ],
+ "score": 1.0,
+ "content": "competence with APPS. In Advances in Neural Information Processing Systems (NeurIPS), 2021a.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13,
+ "bbox_fs": [
+ 106,
+ 240,
+ 506,
+ 275
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 282,
+ 504,
+ 316
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 282,
+ 505,
+ 295
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 294,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 294,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "Steinhardt. Measuring massive multitask language understanding. In International Conference on",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 116,
+ 305,
+ 284,
+ 317
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 305,
+ 284,
+ 317
+ ],
+ "score": 1.0,
+ "content": "Learning Representations (ICLR), 2021b.",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ }
+ ],
+ "index": 16,
+ "bbox_fs": [
+ 106,
+ 282,
+ 505,
+ 317
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 324,
+ 506,
+ 358
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 323,
+ 506,
+ 338
+ ],
+ "score": 1.0,
+ "content": "Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song,",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 335,
+ 505,
+ 348
+ ],
+ "score": 1.0,
+ "content": "and Jacob Steinhardt. Measuring mathematical problem solving with the MATH dataset. In",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 347,
+ 403,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 347,
+ 403,
+ 359
+ ],
+ "score": 1.0,
+ "content": "Advances in Neural Information Processing Systems (NeurIPS), 2021c.",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ }
+ ],
+ "index": 19,
+ "bbox_fs": [
+ 105,
+ 323,
+ 506,
+ 359
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 366,
+ 505,
+ 390
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 365,
+ 506,
+ 381
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 365,
+ 506,
+ 381
+ ],
+ "score": 1.0,
+ "content": "Abigail Z. Jacobs and Hanna Wallach. Measurement and fairness. In ACM Conference on Fairness,",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 114,
+ 378,
+ 315,
+ 390
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 378,
+ 315,
+ 390
+ ],
+ "score": 1.0,
+ "content": "Accountability, and Transparency (FAccT), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21.5,
+ "bbox_fs": [
+ 105,
+ 365,
+ 506,
+ 390
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 104,
+ 397,
+ 505,
+ 421
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 398,
+ 505,
+ 411
+ ],
+ "score": 1.0,
+ "content": "Karen E. Jacowitz and Daniel Kahneman. Measures of anchoring in estimation tasks. Personality",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 409,
+ 350,
+ 421
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 409,
+ 350,
+ 421
+ ],
+ "score": 1.0,
+ "content": "and Social Psychology Bulletin, 21(11):1161–1166, 1995.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23.5,
+ "bbox_fs": [
+ 106,
+ 398,
+ 505,
+ 421
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 428,
+ 505,
+ 462
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 429,
+ 505,
+ 441
+ ],
+ "score": 1.0,
+ "content": "Daniel Kahneman and Shane Frederick. Representativeness revisited: Attribute substitution in",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 440,
+ 507,
+ 452
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 440,
+ 507,
+ 452
+ ],
+ "score": 1.0,
+ "content": "intuitive judgment. In Heuristics and Biases: The Psychology of Intuitive Judgement, pages 49–81.",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 114,
+ 450,
+ 143,
+ 463
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 450,
+ 143,
+ 463
+ ],
+ "score": 1.0,
+ "content": "2002.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 105,
+ 429,
+ 507,
+ 463
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 470,
+ 504,
+ 494
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 485
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 468,
+ 505,
+ 485
+ ],
+ "score": 1.0,
+ "content": "Kalpesh Krishna, Aurko Roy, and Mohit Iyyer. Hurdles to progress in long-form question answering.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 481,
+ 434,
+ 496
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 481,
+ 434,
+ 496
+ ],
+ "score": 1.0,
+ "content": "In North American Association for Computational Linguistics (NAACL), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 105,
+ 468,
+ 505,
+ 496
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 501,
+ 503,
+ 525
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 501,
+ 505,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Stephanie Lin, Jacob Hilton, and Owain Evans. Truthfulqa: Measuring how models mimic human",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ },
+ {
+ "bbox": [
+ 116,
+ 513,
+ 327,
+ 525
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 513,
+ 327,
+ 525
+ ],
+ "score": 1.0,
+ "content": "falsehoods. arXiv preprint arXiv:2109.07958, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ }
+ ],
+ "index": 30.5,
+ "bbox_fs": [
+ 106,
+ 501,
+ 505,
+ 525
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 532,
+ 505,
+ 556
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 506,
+ 546
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 533,
+ 506,
+ 546
+ ],
+ "score": 1.0,
+ "content": "Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 544,
+ 459,
+ 556
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 544,
+ 459,
+ 556
+ ],
+ "score": 1.0,
+ "content": "makes good in-context examples for GPT-3. arXiv preprint arXiv:2101.06804, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32.5,
+ "bbox_fs": [
+ 106,
+ 533,
+ 506,
+ 556
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 564,
+ 464,
+ 576
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 104,
+ 563,
+ 464,
+ 578
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 104,
+ 563,
+ 464,
+ 578
+ ],
+ "score": 1.0,
+ "content": "David E. Meyer. Semantic priming well established. Science, 345(6196):523–523, 2014.",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 104,
+ 563,
+ 464,
+ 578
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 584,
+ 506,
+ 607
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 597
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 584,
+ 505,
+ 597
+ ],
+ "score": 1.0,
+ "content": "Moin Nadeem, Anna Bethke, and Siva Reddy. Stereoset: Measuring stereotypical bias in pretrained",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 595,
+ 352,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 595,
+ 352,
+ 608
+ ],
+ "score": 1.0,
+ "content": "language models. arXiv preprint arXiv:2004.09456, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35.5,
+ "bbox_fs": [
+ 105,
+ 584,
+ 505,
+ 608
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 615,
+ 505,
+ 649
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 614,
+ 506,
+ 629
+ ],
+ "score": 1.0,
+ "content": "Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huam Wang, Yingbo Zhou, Silvio Savarese,",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 627,
+ 505,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 627,
+ 505,
+ 640
+ ],
+ "score": 1.0,
+ "content": "and Caiming Xiong. A conversational paradigm for program synthesis. arXiv preprint",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ },
+ {
+ "bbox": [
+ 116,
+ 637,
+ 219,
+ 650
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 637,
+ 219,
+ 650
+ ],
+ "score": 1.0,
+ "content": "arXiv:2203.13474, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 105,
+ 614,
+ 506,
+ 650
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 657,
+ 504,
+ 692
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 670
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 658,
+ 505,
+ 670
+ ],
+ "score": 1.0,
+ "content": "Luke Oakden-Rayner, Jared Dunnmon, Gustavo Carneiro, and Christopher Ré. Hidden stratification",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 115,
+ 668,
+ 505,
+ 682
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 668,
+ 505,
+ 682
+ ],
+ "score": 1.0,
+ "content": "causes clinically meaningful failures in machine learning for medical imaging. In Proceedings of",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 115,
+ 679,
+ 438,
+ 693
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 679,
+ 438,
+ 693
+ ],
+ "score": 1.0,
+ "content": "the ACM Conference on Health, Inference, and Learning, pages 151–159, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 105,
+ 658,
+ 505,
+ 693
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 699,
+ 502,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 697,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 116,
+ 711,
+ 398,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 711,
+ 398,
+ 723
+ ],
+ "score": 1.0,
+ "content": "models are unsupervised multitask learners. OpenAI Blog, 1(8), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ }
+ ],
+ "index": 43.5,
+ "bbox_fs": [
+ 105,
+ 697,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 72,
+ 506,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 114,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "J. Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Henni-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 94,
+ 507,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 94,
+ 507,
+ 107
+ ],
+ "score": 1.0,
+ "content": "gan, Jacob Menick, Albin Cassirer, Richard Powell, G. V. D. Driessche, Lisa Anne Hendricks,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 105,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 105,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 115,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 115,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "Huang, Jonathan Uesato, John F. J. Mellor, I. Higgins, Antonia Creswell, Nathan McAleese,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 128,
+ 506,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 128,
+ 506,
+ 139
+ ],
+ "score": 1.0,
+ "content": "Amy Wu, Erich Elsen, Siddhant M. Jayakumar, Elena Buchatskaya, D. Budden, Esme Sutherland,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 137,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 137,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "K. Simonyan, Michela Paganini, L. Sifre, Lena Martens, Xiang Lorraine Li, A. Kuncoro, Aida",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 115,
+ 149,
+ 506,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 149,
+ 506,
+ 162
+ ],
+ "score": 1.0,
+ "content": "Nematzadeh, E. Gribovskaya, Domenic Donato, Angeliki Lazaridou, A. Mensch, J. Lespiau, Maria",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 160,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 160,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Tsimpoukelli, N. Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Tobias Pohlen, Zhitao",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 169,
+ 507,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 169,
+ 507,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Gong, Daniel Toyama, Cyprien de Masson d’Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "I. Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 204
+ ],
+ "score": 1.0,
+ "content": "Matthew Johnson, Blake A. Hechtman, Laura Weidinger, Iason Gabriel, William S. Isaac, Edward",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem W. Ayoub, Jeff",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 116,
+ 215,
+ 507,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 215,
+ 507,
+ 227
+ ],
+ "score": 1.0,
+ "content": "Stanway, L. Bennett, D. Hassabis, K. Kavukcuoglu, and Geoffrey Irving. Scaling language models:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 226,
+ 375,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 226,
+ 375,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Methods, analysis & insights from training gopher. arXiv, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 7
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 243,
+ 506,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 507,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 507,
+ 256
+ ],
+ "score": 1.0,
+ "content": "Ashwin Ram, Rohit Prasad, Chandra Khatri, Anu Venkatesh, Raefer Gabriel, Qing Liu, Jeff Nunn,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 255,
+ 507,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 255,
+ 507,
+ 266
+ ],
+ "score": 1.0,
+ "content": "Behnam Hedayatnia, Ming Cheng, Ashish Nagar, Eric King, Kate Bland, Amanda Wartick, Yi Pan,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 265,
+ 506,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 265,
+ 506,
+ 278
+ ],
+ "score": 1.0,
+ "content": "Han Song, Sk Jayadevan, Gene Hwang, and Art Pettigrue. Conversational ai: The science behind",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 276,
+ 342,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 276,
+ 342,
+ 288
+ ],
+ "score": 1.0,
+ "content": "the alexa prize. arXiv preprint arXiv:1801.03604, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 294,
+ 506,
+ 327
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "Sascha Rothe, Shashi Narayan, and Aliaksei Severyn. Leveraging pre-trained checkpoints for",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 306,
+ 506,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 306,
+ 506,
+ 319
+ ],
+ "score": 1.0,
+ "content": "sequence generation tasks. Transactions of the Association for Computational Linguistics (TACL),",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 116,
+ 317,
+ 189,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 317,
+ 189,
+ 328
+ ],
+ "score": 1.0,
+ "content": "8:264–280, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 506,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 116,
+ 347,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 347,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "neural networks for group shifts: On the importance of regularization for worst-case generalization.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 116,
+ 357,
+ 406,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 357,
+ 406,
+ 369
+ ],
+ "score": 1.0,
+ "content": "In International Conference on Learning Representations (ICLR), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 375,
+ 506,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 388
+ ],
+ "score": 1.0,
+ "content": "Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. The woman worked as",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 385,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 385,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "a babysitter: On biases in language generation. In Empirical Methods in Natural Language",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 397,
+ 231,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 397,
+ 231,
+ 409
+ ],
+ "score": 1.0,
+ "content": "Processing (EMNLP), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 414,
+ 504,
+ 438
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 426,
+ 432,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 426,
+ 432,
+ 438
+ ],
+ "score": 1.0,
+ "content": "reduces hallucination in conversation. arXiv preprint arXiv:2104.07567, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 444,
+ 372,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 373,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 373,
+ 458
+ ],
+ "score": 1.0,
+ "content": "Jacob Steinhardt. Anchor weights for ML. Bounded Regret, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 462,
+ 506,
+ 496
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Dario Amodei, and Paul Christiano. Learning to summarize from human feedback. In Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 485,
+ 347,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 485,
+ 347,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Neural Information Processing Systems (NeurIPS), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 503,
+ 506,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Fritz Strack, Leonard L. Martin, and Nobert Schwarz. Priming and communication: Social determi-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 513,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 513,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": "nants of information use in judgments of life satisfaction. European Journal of Social Psychology,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 525,
+ 205,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 525,
+ 205,
+ 537
+ ],
+ "score": 1.0,
+ "content": "18(5):429–442, 1988.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 542,
+ 504,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "Leonard Tang, Elizabeth Ke, Nikhil Singh, Nakul Verma, and Iddo Drori. Solving probability and",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 554,
+ 447,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 554,
+ 447,
+ 567
+ ],
+ "score": 1.0,
+ "content": "statistics problems by program synthesis. arXiv preprint arXiv:2111.08276, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 572,
+ 506,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 116,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 114,
+ 593,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 593,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 114,
+ 604,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 604,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 615,
+ 507,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 615,
+ 507,
+ 629
+ ],
+ "score": 1.0,
+ "content": "Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen Meier-Hellstern,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 626,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 626,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 637,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 637,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 114,
+ 659,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 659,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 115,
+ 670,
+ 506,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 670,
+ 506,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. LaMDA: Language models for",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 115,
+ 681,
+ 361,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 681,
+ 361,
+ 695
+ ],
+ "score": 1.0,
+ "content": "dialog applications. arXiv preprint arXiv:2201.08239, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 44
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 700,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "Amos Tversky and Daniel Kahneman. Availability: A heuristic for judging frequency and probability.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 115,
+ 711,
+ 292,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 711,
+ 292,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Cognitive Psychology, 5(2):207–232, 1973.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50.5
+ }
+ ],
+ "page_idx": 12,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 312,
+ 754
+ ],
+ "score": 1.0,
+ "content": "13",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 72,
+ 506,
+ 237
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 506,
+ 87
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 506,
+ 87
+ ],
+ "score": 1.0,
+ "content": "Jack W. Rae, Sebastian Borgeaud, Trevor Cai, Katie Millican, Jordan Hoffmann, Francis Song,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 114,
+ 83,
+ 506,
+ 96
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 83,
+ 506,
+ 96
+ ],
+ "score": 1.0,
+ "content": "J. Aslanides, Sarah Henderson, Roman Ring, Susannah Young, Eliza Rutherford, Tom Henni-",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 115,
+ 94,
+ 507,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 94,
+ 507,
+ 107
+ ],
+ "score": 1.0,
+ "content": "gan, Jacob Menick, Albin Cassirer, Richard Powell, G. V. D. Driessche, Lisa Anne Hendricks,",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 105,
+ 506,
+ 118
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 105,
+ 506,
+ 118
+ ],
+ "score": 1.0,
+ "content": "Maribeth Rauh, Po-Sen Huang, Amelia Glaese, Johannes Welbl, Sumanth Dathathri, Saffron",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 115,
+ 115,
+ 506,
+ 129
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 115,
+ 506,
+ 129
+ ],
+ "score": 1.0,
+ "content": "Huang, Jonathan Uesato, John F. J. Mellor, I. Higgins, Antonia Creswell, Nathan McAleese,",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 116,
+ 128,
+ 506,
+ 139
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 128,
+ 506,
+ 139
+ ],
+ "score": 1.0,
+ "content": "Amy Wu, Erich Elsen, Siddhant M. Jayakumar, Elena Buchatskaya, D. Budden, Esme Sutherland,",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 115,
+ 137,
+ 506,
+ 150
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 137,
+ 506,
+ 150
+ ],
+ "score": 1.0,
+ "content": "K. Simonyan, Michela Paganini, L. Sifre, Lena Martens, Xiang Lorraine Li, A. Kuncoro, Aida",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ },
+ {
+ "bbox": [
+ 115,
+ 149,
+ 506,
+ 162
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 149,
+ 506,
+ 162
+ ],
+ "score": 1.0,
+ "content": "Nematzadeh, E. Gribovskaya, Domenic Donato, Angeliki Lazaridou, A. Mensch, J. Lespiau, Maria",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 160,
+ 505,
+ 172
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 160,
+ 505,
+ 172
+ ],
+ "score": 1.0,
+ "content": "Tsimpoukelli, N. Grigorev, Doug Fritz, Thibault Sottiaux, Mantas Pajarskas, Tobias Pohlen, Zhitao",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 115,
+ 169,
+ 507,
+ 184
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 169,
+ 507,
+ 184
+ ],
+ "score": 1.0,
+ "content": "Gong, Daniel Toyama, Cyprien de Masson d’Autume, Yujia Li, Tayfun Terzi, Vladimir Mikulik,",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 506,
+ 195
+ ],
+ "score": 1.0,
+ "content": "I. Babuschkin, Aidan Clark, Diego de Las Casas, Aurelia Guy, Chris Jones, James Bradbury,",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 204
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 193,
+ 506,
+ 204
+ ],
+ "score": 1.0,
+ "content": "Matthew Johnson, Blake A. Hechtman, Laura Weidinger, Iason Gabriel, William S. Isaac, Edward",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 203,
+ 506,
+ 216
+ ],
+ "score": 1.0,
+ "content": "Lockhart, Simon Osindero, Laura Rimell, Chris Dyer, Oriol Vinyals, Kareem W. Ayoub, Jeff",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 116,
+ 215,
+ 507,
+ 227
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 215,
+ 507,
+ 227
+ ],
+ "score": 1.0,
+ "content": "Stanway, L. Bennett, D. Hassabis, K. Kavukcuoglu, and Geoffrey Irving. Scaling language models:",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 116,
+ 226,
+ 375,
+ 238
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 226,
+ 375,
+ 238
+ ],
+ "score": 1.0,
+ "content": "Methods, analysis & insights from training gopher. arXiv, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 7,
+ "bbox_fs": [
+ 105,
+ 71,
+ 507,
+ 238
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 243,
+ 506,
+ 288
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 507,
+ 256
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 243,
+ 507,
+ 256
+ ],
+ "score": 1.0,
+ "content": "Ashwin Ram, Rohit Prasad, Chandra Khatri, Anu Venkatesh, Raefer Gabriel, Qing Liu, Jeff Nunn,",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 116,
+ 255,
+ 507,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 255,
+ 507,
+ 266
+ ],
+ "score": 1.0,
+ "content": "Behnam Hedayatnia, Ming Cheng, Ashish Nagar, Eric King, Kate Bland, Amanda Wartick, Yi Pan,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 265,
+ 506,
+ 278
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 265,
+ 506,
+ 278
+ ],
+ "score": 1.0,
+ "content": "Han Song, Sk Jayadevan, Gene Hwang, and Art Pettigrue. Conversational ai: The science behind",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 276,
+ 342,
+ 288
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 276,
+ 342,
+ 288
+ ],
+ "score": 1.0,
+ "content": "the alexa prize. arXiv preprint arXiv:1801.03604, 2018.",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ }
+ ],
+ "index": 16.5,
+ "bbox_fs": [
+ 105,
+ 243,
+ 507,
+ 288
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 294,
+ 506,
+ 327
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 294,
+ 505,
+ 307
+ ],
+ "score": 1.0,
+ "content": "Sascha Rothe, Shashi Narayan, and Aliaksei Severyn. Leveraging pre-trained checkpoints for",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 115,
+ 306,
+ 506,
+ 319
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 306,
+ 506,
+ 319
+ ],
+ "score": 1.0,
+ "content": "sequence generation tasks. Transactions of the Association for Computational Linguistics (TACL),",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 116,
+ 317,
+ 189,
+ 328
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 317,
+ 189,
+ 328
+ ],
+ "score": 1.0,
+ "content": "8:264–280, 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 20,
+ "bbox_fs": [
+ 106,
+ 294,
+ 506,
+ 328
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 334,
+ 506,
+ 369
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 334,
+ 506,
+ 348
+ ],
+ "score": 1.0,
+ "content": "Shiori Sagawa, Pang Wei Koh, Tatsunori B. Hashimoto, and Percy Liang. Distributionally robust",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ },
+ {
+ "bbox": [
+ 116,
+ 347,
+ 506,
+ 359
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 347,
+ 506,
+ 359
+ ],
+ "score": 1.0,
+ "content": "neural networks for group shifts: On the importance of regularization for worst-case generalization.",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 116,
+ 357,
+ 406,
+ 369
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 357,
+ 406,
+ 369
+ ],
+ "score": 1.0,
+ "content": "In International Conference on Learning Representations (ICLR), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ }
+ ],
+ "index": 23,
+ "bbox_fs": [
+ 106,
+ 334,
+ 506,
+ 369
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 375,
+ 506,
+ 408
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 388
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 374,
+ 506,
+ 388
+ ],
+ "score": 1.0,
+ "content": "Emily Sheng, Kai-Wei Chang, Premkumar Natarajan, and Nanyun Peng. The woman worked as",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ },
+ {
+ "bbox": [
+ 115,
+ 385,
+ 506,
+ 399
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 385,
+ 506,
+ 399
+ ],
+ "score": 1.0,
+ "content": "a babysitter: On biases in language generation. In Empirical Methods in Natural Language",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 116,
+ 397,
+ 231,
+ 409
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 397,
+ 231,
+ 409
+ ],
+ "score": 1.0,
+ "content": "Processing (EMNLP), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ }
+ ],
+ "index": 26,
+ "bbox_fs": [
+ 106,
+ 374,
+ 506,
+ 409
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 414,
+ 504,
+ 438
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 428
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 415,
+ 505,
+ 428
+ ],
+ "score": 1.0,
+ "content": "Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval augmentation",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ },
+ {
+ "bbox": [
+ 115,
+ 426,
+ 432,
+ 438
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 426,
+ 432,
+ 438
+ ],
+ "score": 1.0,
+ "content": "reduces hallucination in conversation. arXiv preprint arXiv:2104.07567, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 28.5,
+ "bbox_fs": [
+ 106,
+ 415,
+ 505,
+ 438
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 444,
+ 372,
+ 456
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 373,
+ 458
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 443,
+ 373,
+ 458
+ ],
+ "score": 1.0,
+ "content": "Jacob Steinhardt. Anchor weights for ML. Bounded Regret, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 105,
+ 443,
+ 373,
+ 458
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 462,
+ 506,
+ 496
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 462,
+ 506,
+ 475
+ ],
+ "score": 1.0,
+ "content": "Nisan Stiennon, Long Ouyang, Jeff Wu, Daniel M. Ziegler, Ryan Lowe, Chelsea Voss, Alec Radford,",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 115,
+ 474,
+ 505,
+ 486
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 474,
+ 505,
+ 486
+ ],
+ "score": 1.0,
+ "content": "Dario Amodei, and Paul Christiano. Learning to summarize from human feedback. In Advances in",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 115,
+ 485,
+ 347,
+ 497
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 485,
+ 347,
+ 497
+ ],
+ "score": 1.0,
+ "content": "Neural Information Processing Systems (NeurIPS), 2020.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ }
+ ],
+ "index": 32,
+ "bbox_fs": [
+ 105,
+ 462,
+ 506,
+ 497
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 503,
+ 506,
+ 536
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 506,
+ 515
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 502,
+ 506,
+ 515
+ ],
+ "score": 1.0,
+ "content": "Fritz Strack, Leonard L. Martin, and Nobert Schwarz. Priming and communication: Social determi-",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 115,
+ 513,
+ 506,
+ 527
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 513,
+ 506,
+ 527
+ ],
+ "score": 1.0,
+ "content": "nants of information use in judgments of life satisfaction. European Journal of Social Psychology,",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 116,
+ 525,
+ 205,
+ 537
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 525,
+ 205,
+ 537
+ ],
+ "score": 1.0,
+ "content": "18(5):429–442, 1988.",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ }
+ ],
+ "index": 35,
+ "bbox_fs": [
+ 106,
+ 502,
+ 506,
+ 537
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 542,
+ 504,
+ 566
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 505,
+ 557
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 542,
+ 505,
+ 557
+ ],
+ "score": 1.0,
+ "content": "Leonard Tang, Elizabeth Ke, Nikhil Singh, Nakul Verma, and Iddo Drori. Solving probability and",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ },
+ {
+ "bbox": [
+ 115,
+ 554,
+ 447,
+ 567
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 554,
+ 447,
+ 567
+ ],
+ "score": 1.0,
+ "content": "statistics problems by program synthesis. arXiv preprint arXiv:2111.08276, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 37.5,
+ "bbox_fs": [
+ 105,
+ 542,
+ 505,
+ 567
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 572,
+ 506,
+ 693
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 585
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 572,
+ 505,
+ 585
+ ],
+ "score": 1.0,
+ "content": "Romal Thoppilan, Daniel De Freitas, Jamie Hall, Noam Shazeer, Apoorv Kulshreshtha, Heng-Tze",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 116,
+ 583,
+ 506,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 583,
+ 506,
+ 596
+ ],
+ "score": 1.0,
+ "content": "Cheng, Alicia Jin, Taylor Bos, Leslie Baker, Yu Du, YaGuang Li, Hongrae Lee, Huaixiu Steven",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ },
+ {
+ "bbox": [
+ 114,
+ 593,
+ 506,
+ 608
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 593,
+ 506,
+ 608
+ ],
+ "score": 1.0,
+ "content": "Zheng, Amin Ghafouri, Marcelo Menegali, Yanping Huang, Maxim Krikun, Dmitry Lepikhin,",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ },
+ {
+ "bbox": [
+ 114,
+ 604,
+ 506,
+ 619
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 604,
+ 506,
+ 619
+ ],
+ "score": 1.0,
+ "content": "James Qin, Dehao Chen, Yuanzhong Xu, Zhifeng Chen, Adam Roberts, Maarten Bosma, Yanqi",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 115,
+ 615,
+ 507,
+ 629
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 615,
+ 507,
+ 629
+ ],
+ "score": 1.0,
+ "content": "Zhou, Chung-Ching Chang, Igor Krivokon, Will Rusch, Marc Pickett, Kathleen Meier-Hellstern,",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 115,
+ 626,
+ 506,
+ 640
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 626,
+ 506,
+ 640
+ ],
+ "score": 1.0,
+ "content": "Meredith Ringel Morris, Tulsee Doshi, Renelito Delos Santos, Toju Duke, Johnny Soraker, Ben",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 115,
+ 637,
+ 506,
+ 651
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 637,
+ 506,
+ 651
+ ],
+ "score": 1.0,
+ "content": "Zevenbergen, Vinodkumar Prabhakaran, Mark Diaz, Ben Hutchinson, Kristen Olson, Alejandra",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 648,
+ 505,
+ 662
+ ],
+ "score": 1.0,
+ "content": "Molina, Erin Hoffman-John, Josh Lee, Lora Aroyo, Ravi Rajakumar, Alena Butryna, Matthew",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 114,
+ 659,
+ 505,
+ 672
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 114,
+ 659,
+ 505,
+ 672
+ ],
+ "score": 1.0,
+ "content": "Lamm, Viktoriya Kuzmina, Joe Fenton, Aaron Cohen, Rachel Bernstein, Ray Kurzweil, Blaise",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ },
+ {
+ "bbox": [
+ 115,
+ 670,
+ 506,
+ 683
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 670,
+ 506,
+ 683
+ ],
+ "score": 1.0,
+ "content": "Aguera-Arcas, Claire Cui, Marian Croak, Ed Chi, and Quoc Le. LaMDA: Language models for",
+ "type": "text"
+ }
+ ],
+ "index": 48
+ },
+ {
+ "bbox": [
+ 115,
+ 681,
+ 361,
+ 695
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 681,
+ 361,
+ 695
+ ],
+ "score": 1.0,
+ "content": "dialog applications. arXiv preprint arXiv:2201.08239, 2022.",
+ "type": "text"
+ }
+ ],
+ "index": 49
+ }
+ ],
+ "index": 44,
+ "bbox_fs": [
+ 107,
+ 572,
+ 507,
+ 695
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 108,
+ 700,
+ 503,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 714
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 698,
+ 505,
+ 714
+ ],
+ "score": 1.0,
+ "content": "Amos Tversky and Daniel Kahneman. Availability: A heuristic for judging frequency and probability.",
+ "type": "text"
+ }
+ ],
+ "index": 50
+ },
+ {
+ "bbox": [
+ 115,
+ 711,
+ 292,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 711,
+ 292,
+ 723
+ ],
+ "score": 1.0,
+ "content": "Cognitive Psychology, 5(2):207–232, 1973.",
+ "type": "text"
+ }
+ ],
+ "index": 51
+ }
+ ],
+ "index": 50.5,
+ "bbox_fs": [
+ 106,
+ 698,
+ 505,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "Amos Tversky and Daniel Kahneman. Judgment under uncertainty: Heuristics and biases. Science,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 84,
+ 235,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 84,
+ 235,
+ 95
+ ],
+ "score": 1.0,
+ "content": "185(4157):1124–1131, 1974.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 101,
+ 505,
+ 124
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 115
+ ],
+ "score": 1.0,
+ "content": "Amos Tversky and Daniel Kahneman. The framing of decisions and the psychology of choice.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 113,
+ 262,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 113,
+ 262,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Science, 211(4481):453–458, 1981.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 131,
+ 504,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 144
+ ],
+ "score": 1.0,
+ "content": "Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 141,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 141,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "triggers for attacking and analyzing NLP. In Empirical Methods in Natural Language Processing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 116,
+ 153,
+ 185,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 153,
+ 185,
+ 165
+ ],
+ "score": 1.0,
+ "content": "(EMNLP), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 171,
+ 505,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 505,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 505,
+ 196
+ ],
+ "score": 1.0,
+ "content": "Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-purpose language",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 194,
+ 506,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 194,
+ 506,
+ 206
+ ],
+ "score": 1.0,
+ "content": "understanding systems. In Advances in Neural Information Processing Systems (NeurIPS), 2019a.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 211,
+ 504,
+ 245
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 211,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 211,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Alex Wang, Amapreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. GLUE:",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 116,
+ 222,
+ 505,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 222,
+ 505,
+ 235
+ ],
+ "score": 1.0,
+ "content": "A multi-task benchmark and analysis platform for natural language understanding. In International",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 232,
+ 345,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 232,
+ 345,
+ 246
+ ],
+ "score": 1.0,
+ "content": "Conference on Learning Representations (ICLR), 2019b.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 251,
+ 505,
+ 274
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 251,
+ 506,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 251,
+ 506,
+ 264
+ ],
+ "score": 1.0,
+ "content": "Ben Wang and Aran Komatsuzaki. GPT-J-6B: A 6 billion parameter autoregressive language model,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 261,
+ 144,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 261,
+ 144,
+ 275
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 281,
+ 506,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 302,
+ 507,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 302,
+ 507,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "William Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. Ethical and social risks of",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 325,
+ 397,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 325,
+ 397,
+ 338
+ ],
+ "score": 1.0,
+ "content": "harm from language models. arXiv preprint arXiv:2112.04359, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 343,
+ 505,
+ 377
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 507,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 507,
+ 356
+ ],
+ "score": 1.0,
+ "content": "Sonja Windhager, Florian Hutzler, Claus-Christian Carbon, Elisabeth Oberzaucher, Katrin Schaefer,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 353,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 353,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "Truls Thorstensen, Helmut Leder, and Karl Grammer. Laying eyes on headlights: Eye movements",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 365,
+ 452,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 365,
+ 452,
+ 378
+ ],
+ "score": 1.0,
+ "content": "suggest facial features in cars. Collegium Antropologicum, 34(3):1075–1080, 2010.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 383,
+ 505,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 382,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 382,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 394,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 394,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "few-shot performance of language models. In International Conference on Machine Learning",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 405,
+ 176,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 405,
+ 176,
+ 418
+ ],
+ "score": 1.0,
+ "content": "(ICML), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 423,
+ 505,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 117,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 447,
+ 254,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 447,
+ 254,
+ 457
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1909.08593, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 477,
+ 156,
+ 490
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 158,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 158,
+ 492
+ ],
+ "score": 1.0,
+ "content": "Checklist",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 498,
+ 208,
+ 509
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 496,
+ 210,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 496,
+ 210,
+ 510
+ ],
+ "score": 1.0,
+ "content": "1. For all authors...",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 513,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 512,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 512,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
+ "type": "text"
+ }
+ ],
+ "index": 31
+ },
+ {
+ "bbox": [
+ 162,
+ 523,
+ 287,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 523,
+ 287,
+ 535
+ ],
+ "score": 1.0,
+ "content": "contributions and scope? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 32
+ },
+ {
+ "bbox": [
+ 145,
+ 536,
+ 440,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 536,
+ 440,
+ 547
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe the limitations of your work? [Yes] See discussion.",
+ "type": "text"
+ }
+ ],
+ "index": 33
+ },
+ {
+ "bbox": [
+ 146,
+ 547,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 547,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] See",
+ "type": "text"
+ }
+ ],
+ "index": 34
+ },
+ {
+ "bbox": [
+ 161,
+ 559,
+ 209,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 559,
+ 209,
+ 572
+ ],
+ "score": 1.0,
+ "content": "discussion.",
+ "type": "text"
+ }
+ ],
+ "index": 35
+ },
+ {
+ "bbox": [
+ 146,
+ 570,
+ 505,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 570,
+ 505,
+ 586
+ ],
+ "score": 1.0,
+ "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
+ "type": "text"
+ }
+ ],
+ "index": 36
+ },
+ {
+ "bbox": [
+ 162,
+ 582,
+ 214,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 582,
+ 214,
+ 596
+ ],
+ "score": 1.0,
+ "content": "them? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 37
+ }
+ ],
+ "index": 34
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 598,
+ 302,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 597,
+ 304,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 597,
+ 304,
+ 611
+ ],
+ "score": 1.0,
+ "content": "2. If you are including theoretical results...",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 612,
+ 452,
+ 637
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 612,
+ 452,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 612,
+ 452,
+ 625
+ ],
+ "score": 1.0,
+ "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 145,
+ 624,
+ 424,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 624,
+ 424,
+ 638
+ ],
+ "score": 1.0,
+ "content": "(b) Did you include complete proofs of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 640,
+ 241,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 638,
+ 243,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 638,
+ 243,
+ 654
+ ],
+ "score": 1.0,
+ "content": "3. If you ran experiments...",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 655,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 654,
+ 507,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 654,
+ 507,
+ 667
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-",
+ "type": "text"
+ }
+ ],
+ "index": 42
+ },
+ {
+ "bbox": [
+ 162,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "mental results (either in the supplemental material or as a URL)? [Yes] Included in the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 162,
+ 677,
+ 212,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 677,
+ 212,
+ 689
+ ],
+ "score": 1.0,
+ "content": "supplement.",
+ "type": "text"
+ }
+ ],
+ "index": 44
+ },
+ {
+ "bbox": [
+ 145,
+ 688,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 688,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they",
+ "type": "text"
+ }
+ ],
+ "index": 45
+ },
+ {
+ "bbox": [
+ 160,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "were chosen)? [Yes] We do not train new models, but in Section 3.1 and Section 4 we",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 160,
+ 711,
+ 346,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 711,
+ 346,
+ 723
+ ],
+ "score": 1.0,
+ "content": "provide details about the models we evaluate.",
+ "type": "text"
+ }
+ ],
+ "index": 47
+ }
+ ],
+ "index": 44.5
+ }
+ ],
+ "page_idx": 13,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "14",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 72,
+ 505,
+ 95
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 71,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "Amos Tversky and Daniel Kahneman. Judgment under uncertainty: Heuristics and biases. Science,",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 116,
+ 84,
+ 235,
+ 95
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 84,
+ 235,
+ 95
+ ],
+ "score": 1.0,
+ "content": "185(4157):1124–1131, 1974.",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ }
+ ],
+ "index": 0.5,
+ "bbox_fs": [
+ 105,
+ 71,
+ 507,
+ 95
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 101,
+ 505,
+ 124
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 115
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 100,
+ 506,
+ 115
+ ],
+ "score": 1.0,
+ "content": "Amos Tversky and Daniel Kahneman. The framing of decisions and the psychology of choice.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 115,
+ 113,
+ 262,
+ 125
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 113,
+ 262,
+ 125
+ ],
+ "score": 1.0,
+ "content": "Science, 211(4481):453–458, 1981.",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ }
+ ],
+ "index": 2.5,
+ "bbox_fs": [
+ 105,
+ 100,
+ 506,
+ 125
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 131,
+ 504,
+ 164
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 144
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 131,
+ 505,
+ 144
+ ],
+ "score": 1.0,
+ "content": "Eric Wallace, Shi Feng, Nikhil Kandpal, Matt Gardner, and Sameer Singh. Universal adversarial",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 115,
+ 141,
+ 506,
+ 156
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 141,
+ 506,
+ 156
+ ],
+ "score": 1.0,
+ "content": "triggers for attacking and analyzing NLP. In Empirical Methods in Natural Language Processing",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ },
+ {
+ "bbox": [
+ 116,
+ 153,
+ 185,
+ 165
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 153,
+ 185,
+ 165
+ ],
+ "score": 1.0,
+ "content": "(EMNLP), 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 5,
+ "bbox_fs": [
+ 106,
+ 131,
+ 506,
+ 165
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 106,
+ 171,
+ 505,
+ 205
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 506,
+ 185
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 171,
+ 506,
+ 185
+ ],
+ "score": 1.0,
+ "content": "Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 115,
+ 181,
+ 505,
+ 196
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 181,
+ 505,
+ 196
+ ],
+ "score": 1.0,
+ "content": "Levy, and Samuel R. Bowman. SuperGLUE: A stickier benchmark for general-purpose language",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 116,
+ 194,
+ 506,
+ 206
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 194,
+ 506,
+ 206
+ ],
+ "score": 1.0,
+ "content": "understanding systems. In Advances in Neural Information Processing Systems (NeurIPS), 2019a.",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ }
+ ],
+ "index": 8,
+ "bbox_fs": [
+ 106,
+ 171,
+ 506,
+ 206
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 109,
+ 211,
+ 504,
+ 245
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 107,
+ 211,
+ 506,
+ 225
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 107,
+ 211,
+ 506,
+ 225
+ ],
+ "score": 1.0,
+ "content": "Alex Wang, Amapreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. GLUE:",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 116,
+ 222,
+ 505,
+ 235
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 116,
+ 222,
+ 505,
+ 235
+ ],
+ "score": 1.0,
+ "content": "A multi-task benchmark and analysis platform for natural language understanding. In International",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 115,
+ 232,
+ 345,
+ 246
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 232,
+ 345,
+ 246
+ ],
+ "score": 1.0,
+ "content": "Conference on Learning Representations (ICLR), 2019b.",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ }
+ ],
+ "index": 11,
+ "bbox_fs": [
+ 107,
+ 211,
+ 506,
+ 246
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 105,
+ 251,
+ 505,
+ 274
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 251,
+ 506,
+ 264
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 251,
+ 506,
+ 264
+ ],
+ "score": 1.0,
+ "content": "Ben Wang and Aran Komatsuzaki. GPT-J-6B: A 6 billion parameter autoregressive language model,",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 115,
+ 261,
+ 144,
+ 275
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 261,
+ 144,
+ 275
+ ],
+ "score": 1.0,
+ "content": "2021.",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 13.5,
+ "bbox_fs": [
+ 105,
+ 251,
+ 506,
+ 275
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 281,
+ 506,
+ 337
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 505,
+ 293
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 281,
+ 505,
+ 293
+ ],
+ "score": 1.0,
+ "content": "Laura Weidinger, John Mellor, Maribeth Rauh, Conor Griffin, Jonathan Uesato, Po-Sen Huang, Myra",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ },
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 306
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 291,
+ 506,
+ 306
+ ],
+ "score": 1.0,
+ "content": "Cheng, Mia Glaese, Borja Balle, Atoosa Kasirzadeh, Zac Kenton, Sasha Brown, Will Hawkins,",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 115,
+ 302,
+ 507,
+ 316
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 302,
+ 507,
+ 316
+ ],
+ "score": 1.0,
+ "content": "Tom Stepleton, Courtney Biles, Abeba Birhane, Julia Haas, Laura Rimell, Lisa Anne Hendricks,",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 326
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 314,
+ 505,
+ 326
+ ],
+ "score": 1.0,
+ "content": "William Isaac, Sean Legassick, Geoffrey Irving, and Iason Gabriel. Ethical and social risks of",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 115,
+ 325,
+ 397,
+ 338
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 325,
+ 397,
+ 338
+ ],
+ "score": 1.0,
+ "content": "harm from language models. arXiv preprint arXiv:2112.04359, 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ }
+ ],
+ "index": 17,
+ "bbox_fs": [
+ 106,
+ 281,
+ 507,
+ 338
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 343,
+ 505,
+ 377
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 507,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 343,
+ 507,
+ 356
+ ],
+ "score": 1.0,
+ "content": "Sonja Windhager, Florian Hutzler, Claus-Christian Carbon, Elisabeth Oberzaucher, Katrin Schaefer,",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 115,
+ 353,
+ 505,
+ 367
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 353,
+ 505,
+ 367
+ ],
+ "score": 1.0,
+ "content": "Truls Thorstensen, Helmut Leder, and Karl Grammer. Laying eyes on headlights: Eye movements",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ },
+ {
+ "bbox": [
+ 115,
+ 365,
+ 452,
+ 378
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 365,
+ 452,
+ 378
+ ],
+ "score": 1.0,
+ "content": "suggest facial features in cars. Collegium Antropologicum, 34(3):1075–1080, 2010.",
+ "type": "text"
+ }
+ ],
+ "index": 22
+ }
+ ],
+ "index": 21,
+ "bbox_fs": [
+ 105,
+ 343,
+ 507,
+ 378
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 383,
+ 505,
+ 417
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 382,
+ 505,
+ 397
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 382,
+ 505,
+ 397
+ ],
+ "score": 1.0,
+ "content": "Tony Z. Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving",
+ "type": "text"
+ }
+ ],
+ "index": 23
+ },
+ {
+ "bbox": [
+ 115,
+ 394,
+ 505,
+ 408
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 394,
+ 505,
+ 408
+ ],
+ "score": 1.0,
+ "content": "few-shot performance of language models. In International Conference on Machine Learning",
+ "type": "text"
+ }
+ ],
+ "index": 24
+ },
+ {
+ "bbox": [
+ 115,
+ 405,
+ 176,
+ 418
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 405,
+ 176,
+ 418
+ ],
+ "score": 1.0,
+ "content": "(ICML), 2021.",
+ "type": "text"
+ }
+ ],
+ "index": 25
+ }
+ ],
+ "index": 24,
+ "bbox_fs": [
+ 105,
+ 382,
+ 505,
+ 418
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 107,
+ 423,
+ 505,
+ 457
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 436
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 106,
+ 424,
+ 505,
+ 436
+ ],
+ "score": 1.0,
+ "content": "Daniel M. Ziegler, Nisan Stiennon, Jeffrey Wu, Tom B. Brown, Alec Radford, Dario Amodei, Paul",
+ "type": "text"
+ }
+ ],
+ "index": 26
+ },
+ {
+ "bbox": [
+ 117,
+ 435,
+ 505,
+ 447
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 117,
+ 435,
+ 505,
+ 447
+ ],
+ "score": 1.0,
+ "content": "Christiano, and Geoffrey Irving. Fine-tuning language models from human preferences. arXiv",
+ "type": "text"
+ }
+ ],
+ "index": 27
+ },
+ {
+ "bbox": [
+ 115,
+ 447,
+ 254,
+ 457
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 115,
+ 447,
+ 254,
+ 457
+ ],
+ "score": 1.0,
+ "content": "preprint arXiv:1909.08593, 2019.",
+ "type": "text"
+ }
+ ],
+ "index": 28
+ }
+ ],
+ "index": 27,
+ "bbox_fs": [
+ 106,
+ 424,
+ 505,
+ 457
+ ]
+ },
+ {
+ "type": "title",
+ "bbox": [
+ 107,
+ 477,
+ 156,
+ 490
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 158,
+ 492
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 105,
+ 476,
+ 158,
+ 492
+ ],
+ "score": 1.0,
+ "content": "Checklist",
+ "type": "text"
+ }
+ ],
+ "index": 29
+ }
+ ],
+ "index": 29
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 498,
+ 208,
+ 509
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 496,
+ 210,
+ 510
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 496,
+ 210,
+ 510
+ ],
+ "score": 1.0,
+ "content": "1. For all authors...",
+ "type": "text"
+ }
+ ],
+ "index": 30
+ }
+ ],
+ "index": 30,
+ "bbox_fs": [
+ 129,
+ 496,
+ 210,
+ 510
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 513,
+ 505,
+ 595
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 146,
+ 512,
+ 505,
+ 524
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 512,
+ 505,
+ 524
+ ],
+ "score": 1.0,
+ "content": "(a) Do the main claims made in the abstract and introduction accurately reflect the paper’s",
+ "type": "text"
+ }
+ ],
+ "index": 31,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 523,
+ 287,
+ 535
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 523,
+ 287,
+ 535
+ ],
+ "score": 1.0,
+ "content": "contributions and scope? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 32,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 536,
+ 440,
+ 547
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 536,
+ 440,
+ 547
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe the limitations of your work? [Yes] See discussion.",
+ "type": "text"
+ }
+ ],
+ "index": 33,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 547,
+ 505,
+ 562
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 547,
+ 505,
+ 562
+ ],
+ "score": 1.0,
+ "content": "(c) Did you discuss any potential negative societal impacts of your work? [Yes] See",
+ "type": "text"
+ }
+ ],
+ "index": 34,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 559,
+ 209,
+ 572
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 559,
+ 209,
+ 572
+ ],
+ "score": 1.0,
+ "content": "discussion.",
+ "type": "text"
+ }
+ ],
+ "index": 35,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 570,
+ 505,
+ 586
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 570,
+ 505,
+ 586
+ ],
+ "score": 1.0,
+ "content": "(d) Have you read the ethics review guidelines and ensured that your paper conforms to",
+ "type": "text"
+ }
+ ],
+ "index": 36,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 582,
+ 214,
+ 596
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 582,
+ 214,
+ 596
+ ],
+ "score": 1.0,
+ "content": "them? [Yes]",
+ "type": "text"
+ }
+ ],
+ "index": 37,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 34,
+ "bbox_fs": [
+ 145,
+ 512,
+ 505,
+ 596
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 598,
+ 302,
+ 609
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 129,
+ 597,
+ 304,
+ 611
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 129,
+ 597,
+ 304,
+ 611
+ ],
+ "score": 1.0,
+ "content": "2. If you are including theoretical results...",
+ "type": "text"
+ }
+ ],
+ "index": 38
+ }
+ ],
+ "index": 38,
+ "bbox_fs": [
+ 129,
+ 597,
+ 304,
+ 611
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 612,
+ 452,
+ 637
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 612,
+ 452,
+ 625
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 612,
+ 452,
+ 625
+ ],
+ "score": 1.0,
+ "content": "(a) Did you state the full set of assumptions of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 39
+ },
+ {
+ "bbox": [
+ 145,
+ 624,
+ 424,
+ 638
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 624,
+ 424,
+ 638
+ ],
+ "score": 1.0,
+ "content": "(b) Did you include complete proofs of all theoretical results? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 40
+ }
+ ],
+ "index": 39.5,
+ "bbox_fs": [
+ 145,
+ 612,
+ 452,
+ 638
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 640,
+ 241,
+ 651
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 638,
+ 243,
+ 654
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 638,
+ 243,
+ 654
+ ],
+ "score": 1.0,
+ "content": "3. If you ran experiments...",
+ "type": "text"
+ }
+ ],
+ "index": 41
+ }
+ ],
+ "index": 41,
+ "bbox_fs": [
+ 128,
+ 638,
+ 243,
+ 654
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 655,
+ 505,
+ 722
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 654,
+ 507,
+ 667
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 654,
+ 507,
+ 667
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the code, data, and instructions needed to reproduce the main experi-",
+ "type": "text"
+ }
+ ],
+ "index": 42,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 665,
+ 505,
+ 678
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 665,
+ 505,
+ 678
+ ],
+ "score": 1.0,
+ "content": "mental results (either in the supplemental material or as a URL)? [Yes] Included in the",
+ "type": "text"
+ }
+ ],
+ "index": 43
+ },
+ {
+ "bbox": [
+ 162,
+ 677,
+ 212,
+ 689
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 677,
+ 212,
+ 689
+ ],
+ "score": 1.0,
+ "content": "supplement.",
+ "type": "text"
+ }
+ ],
+ "index": 44,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 688,
+ 505,
+ 702
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 688,
+ 505,
+ 702
+ ],
+ "score": 1.0,
+ "content": "(b) Did you specify all the training details (e.g., data splits, hyperparameters, how they",
+ "type": "text"
+ }
+ ],
+ "index": 45,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 160,
+ 699,
+ 506,
+ 713
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 699,
+ 506,
+ 713
+ ],
+ "score": 1.0,
+ "content": "were chosen)? [Yes] We do not train new models, but in Section 3.1 and Section 4 we",
+ "type": "text"
+ }
+ ],
+ "index": 46
+ },
+ {
+ "bbox": [
+ 160,
+ 711,
+ 346,
+ 723
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 160,
+ 711,
+ 346,
+ 723
+ ],
+ "score": 1.0,
+ "content": "provide details about the models we evaluate.",
+ "type": "text"
+ }
+ ],
+ "index": 47,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 72,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 0,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 83,
+ 507,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 83,
+ 507,
+ 97
+ ],
+ "score": 1.0,
+ "content": "ments multiple times)? [N/A] Our experiments consider the deterministic rollout of",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 161,
+ 94,
+ 263,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 94,
+ 263,
+ 107
+ ],
+ "score": 1.0,
+ "content": "models on fixed prompts.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 2,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 107,
+ 506,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 107,
+ 506,
+ 120
+ ],
+ "score": 1.0,
+ "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 3,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 117,
+ 506,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 117,
+ 506,
+ 130
+ ],
+ "score": 1.0,
+ "content": "of GPUs, internal cluster, or cloud provider)? [N/A] We do not train models in this",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 161,
+ 129,
+ 187,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 129,
+ 187,
+ 141
+ ],
+ "score": 1.0,
+ "content": "work.",
+ "type": "text",
+ "cross_page": true
+ }
+ ],
+ "index": 5,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 44.5,
+ "bbox_fs": [
+ 145,
+ 654,
+ 507,
+ 723
+ ]
+ }
+ ]
+ },
+ {
+ "preproc_blocks": [
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 73,
+ 506,
+ 140
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 507,
+ 86
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 72,
+ 507,
+ 86
+ ],
+ "score": 1.0,
+ "content": "(c) Did you report error bars (e.g., with respect to the random seed after running experi-",
+ "type": "text"
+ }
+ ],
+ "index": 0
+ },
+ {
+ "bbox": [
+ 161,
+ 83,
+ 507,
+ 97
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 83,
+ 507,
+ 97
+ ],
+ "score": 1.0,
+ "content": "ments multiple times)? [N/A] Our experiments consider the deterministic rollout of",
+ "type": "text"
+ }
+ ],
+ "index": 1
+ },
+ {
+ "bbox": [
+ 161,
+ 94,
+ 263,
+ 107
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 94,
+ 263,
+ 107
+ ],
+ "score": 1.0,
+ "content": "models on fixed prompts.",
+ "type": "text"
+ }
+ ],
+ "index": 2
+ },
+ {
+ "bbox": [
+ 146,
+ 107,
+ 506,
+ 120
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 107,
+ 506,
+ 120
+ ],
+ "score": 1.0,
+ "content": "(d) Did you include the total amount of compute and the type of resources used (e.g., type",
+ "type": "text"
+ }
+ ],
+ "index": 3
+ },
+ {
+ "bbox": [
+ 161,
+ 117,
+ 506,
+ 130
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 117,
+ 506,
+ 130
+ ],
+ "score": 1.0,
+ "content": "of GPUs, internal cluster, or cloud provider)? [N/A] We do not train models in this",
+ "type": "text"
+ }
+ ],
+ "index": 4
+ },
+ {
+ "bbox": [
+ 161,
+ 129,
+ 187,
+ 141
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 129,
+ 187,
+ 141
+ ],
+ "score": 1.0,
+ "content": "work.",
+ "type": "text"
+ }
+ ],
+ "index": 5
+ }
+ ],
+ "index": 2.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 144,
+ 504,
+ 156
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 144,
+ 506,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 144,
+ 506,
+ 158
+ ],
+ "score": 1.0,
+ "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 144,
+ 158,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 158,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 158,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] We cite GPT-3 and",
+ "type": "text"
+ }
+ ],
+ "index": 7
+ },
+ {
+ "bbox": [
+ 162,
+ 169,
+ 299,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 169,
+ 299,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Codex, and give credit to OpenAI.",
+ "type": "text"
+ }
+ ],
+ "index": 8
+ },
+ {
+ "bbox": [
+ 145,
+ 183,
+ 360,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 183,
+ 360,
+ 195
+ ],
+ "score": 1.0,
+ "content": "(b) Did you mention the license of the assets? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 9
+ },
+ {
+ "bbox": [
+ 145,
+ 195,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 195,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 10
+ },
+ {
+ "bbox": [
+ 145,
+ 219,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 219,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
+ "type": "text"
+ }
+ ],
+ "index": 11
+ },
+ {
+ "bbox": [
+ 162,
+ 231,
+ 253,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 231,
+ 253,
+ 243
+ ],
+ "score": 1.0,
+ "content": "using/curating? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 12
+ },
+ {
+ "bbox": [
+ 146,
+ 242,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 242,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
+ "type": "text"
+ }
+ ],
+ "index": 13
+ },
+ {
+ "bbox": [
+ 162,
+ 254,
+ 325,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 254,
+ 325,
+ 266
+ ],
+ "score": 1.0,
+ "content": "information or offensive content? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 14
+ }
+ ],
+ "index": 10.5
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 269,
+ 432,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 267,
+ 433,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 267,
+ 433,
+ 283
+ ],
+ "score": 1.0,
+ "content": "5. If you used crowdsourcing or conducted research with human subjects...",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 146,
+ 284,
+ 505,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 283,
+ 506,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 283,
+ 506,
+ 297
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
+ "type": "text"
+ }
+ ],
+ "index": 16
+ },
+ {
+ "bbox": [
+ 161,
+ 295,
+ 237,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 295,
+ 237,
+ 307
+ ],
+ "score": 1.0,
+ "content": "applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 17
+ },
+ {
+ "bbox": [
+ 145,
+ 308,
+ 505,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 308,
+ 505,
+ 320
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
+ "type": "text"
+ }
+ ],
+ "index": 18
+ },
+ {
+ "bbox": [
+ 161,
+ 318,
+ 342,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 318,
+ 342,
+ 331
+ ],
+ "score": 1.0,
+ "content": "Board (IRB) approvals, if applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 19
+ },
+ {
+ "bbox": [
+ 147,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
+ "type": "text"
+ }
+ ],
+ "index": 20
+ },
+ {
+ "bbox": [
+ 162,
+ 343,
+ 333,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 343,
+ 333,
+ 356
+ ],
+ "score": 1.0,
+ "content": "spent on participant compensation? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 21
+ }
+ ],
+ "index": 18.5
+ }
+ ],
+ "page_idx": 14,
+ "page_size": [
+ 612,
+ 792
+ ],
+ "discarded_blocks": [
+ {
+ "type": "discarded",
+ "bbox": [
+ 300,
+ 741,
+ 311,
+ 750
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 299,
+ 740,
+ 313,
+ 754
+ ],
+ "score": 1.0,
+ "content": "15",
+ "type": "text"
+ }
+ ]
+ }
+ ]
+ }
+ ],
+ "para_blocks": [
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 73,
+ 506,
+ 140
+ ],
+ "lines": [],
+ "index": 2.5,
+ "bbox_fs": [
+ 145,
+ 72,
+ 507,
+ 141
+ ],
+ "lines_deleted": true
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 130,
+ 144,
+ 504,
+ 156
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 144,
+ 506,
+ 158
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 144,
+ 506,
+ 158
+ ],
+ "score": 1.0,
+ "content": "4. If you are using existing assets (e.g., code, data, models) or curating/releasing new assets...",
+ "type": "text"
+ }
+ ],
+ "index": 6
+ }
+ ],
+ "index": 6,
+ "bbox_fs": [
+ 128,
+ 144,
+ 506,
+ 158
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 144,
+ 158,
+ 505,
+ 266
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 158,
+ 506,
+ 171
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 158,
+ 506,
+ 171
+ ],
+ "score": 1.0,
+ "content": "(a) If your work uses existing assets, did you cite the creators? [Yes] We cite GPT-3 and",
+ "type": "text"
+ }
+ ],
+ "index": 7,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 169,
+ 299,
+ 182
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 169,
+ 299,
+ 182
+ ],
+ "score": 1.0,
+ "content": "Codex, and give credit to OpenAI.",
+ "type": "text"
+ }
+ ],
+ "index": 8,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 183,
+ 360,
+ 195
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 183,
+ 360,
+ 195
+ ],
+ "score": 1.0,
+ "content": "(b) Did you mention the license of the assets? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 9,
+ "is_list_start_line": true,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 195,
+ 506,
+ 208
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 195,
+ 506,
+ 208
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include any new assets either in the supplemental material or as a URL? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 10,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 219,
+ 505,
+ 233
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 219,
+ 505,
+ 233
+ ],
+ "score": 1.0,
+ "content": "(d) Did you discuss whether and how consent was obtained from people whose data you’re",
+ "type": "text"
+ }
+ ],
+ "index": 11,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 231,
+ 253,
+ 243
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 231,
+ 253,
+ 243
+ ],
+ "score": 1.0,
+ "content": "using/curating? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 12,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 146,
+ 242,
+ 505,
+ 257
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 146,
+ 242,
+ 505,
+ 257
+ ],
+ "score": 1.0,
+ "content": "(e) Did you discuss whether the data you are using/curating contains personally identifiable",
+ "type": "text"
+ }
+ ],
+ "index": 13,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 254,
+ 325,
+ 266
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 254,
+ 325,
+ 266
+ ],
+ "score": 1.0,
+ "content": "information or offensive content? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 14,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 10.5,
+ "bbox_fs": [
+ 145,
+ 158,
+ 506,
+ 266
+ ]
+ },
+ {
+ "type": "text",
+ "bbox": [
+ 131,
+ 269,
+ 432,
+ 281
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 128,
+ 267,
+ 433,
+ 283
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 128,
+ 267,
+ 433,
+ 283
+ ],
+ "score": 1.0,
+ "content": "5. If you used crowdsourcing or conducted research with human subjects...",
+ "type": "text"
+ }
+ ],
+ "index": 15
+ }
+ ],
+ "index": 15,
+ "bbox_fs": [
+ 128,
+ 267,
+ 433,
+ 283
+ ]
+ },
+ {
+ "type": "list",
+ "bbox": [
+ 146,
+ 284,
+ 505,
+ 354
+ ],
+ "lines": [
+ {
+ "bbox": [
+ 145,
+ 283,
+ 506,
+ 297
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 283,
+ 506,
+ 297
+ ],
+ "score": 1.0,
+ "content": "(a) Did you include the full text of instructions given to participants and screenshots, if",
+ "type": "text"
+ }
+ ],
+ "index": 16,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 295,
+ 237,
+ 307
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 295,
+ 237,
+ 307
+ ],
+ "score": 1.0,
+ "content": "applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 17,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 145,
+ 308,
+ 505,
+ 320
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 145,
+ 308,
+ 505,
+ 320
+ ],
+ "score": 1.0,
+ "content": "(b) Did you describe any potential participant risks, with links to Institutional Review",
+ "type": "text"
+ }
+ ],
+ "index": 18,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 161,
+ 318,
+ 342,
+ 331
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 161,
+ 318,
+ 342,
+ 331
+ ],
+ "score": 1.0,
+ "content": "Board (IRB) approvals, if applicable? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 19,
+ "is_list_end_line": true
+ },
+ {
+ "bbox": [
+ 147,
+ 331,
+ 505,
+ 344
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 147,
+ 331,
+ 505,
+ 344
+ ],
+ "score": 1.0,
+ "content": "(c) Did you include the estimated hourly wage paid to participants and the total amount",
+ "type": "text"
+ }
+ ],
+ "index": 20,
+ "is_list_start_line": true
+ },
+ {
+ "bbox": [
+ 162,
+ 343,
+ 333,
+ 356
+ ],
+ "spans": [
+ {
+ "bbox": [
+ 162,
+ 343,
+ 333,
+ 356
+ ],
+ "score": 1.0,
+ "content": "spent on participant compensation? [N/A]",
+ "type": "text"
+ }
+ ],
+ "index": 21,
+ "is_list_end_line": true
+ }
+ ],
+ "index": 18.5,
+ "bbox_fs": [
+ 145,
+ 283,
+ 506,
+ 356
+ ]
+ }
+ ]
+ }
+ ],
+ "_backend": "pipeline",
+ "_version_name": "2.2.2"
+}
\ No newline at end of file