README.md

metadata

license: apache-2.0
task_categories:
  - visual-question-answering
  - image-text-to-text
language:
  - en
tags:
  - multimodal
  - benchmark
  - vision-language
  - mllm
  - counter-intuitive
  - commonsense-reasoning
  - language-bias
size_categories:
  - n<1K
pretty_name: 'CAIT: Counter-intuitive Action Image Test'

CAIT: Counter-intuitive Action Image Test

CAIT is an evaluation benchmark of 400 high-fidelity synthetic scenes in which the visual evidence deliberately contradicts everyday common sense, for example "a rabbit is chasing a tiger". Each item is a binary forced-choice question between the scene that is actually depicted and the commonsense-consistent scene that is not. Models that lean on language priors instead of looking at the image pick the plausible-sounding option and fail.

The benchmark accompanies the paper Seeing vs. Believing: Evaluating the Language Bias of Open-Source MLLMs in Counter-Intuitive Scenes. Humans reach about 0.95 accuracy on it and leading proprietary models up to about 0.88, while standard open-source instruction-tuned models perform at chance level.

This is an evaluation-only benchmark; it is published as a single test split and is not intended for training.

Loading

from datasets import load_dataset

ds = load_dataset("LukeLing/CAIT", split="test")
print(ds)
print(ds[0]["question"], ds[0]["option_a"], ds[0]["option_b"], ds[0]["answer"])
ds[0]["image"]  # a PIL image

Fields

Field	Type	Description
image	image	The scene, resolved by the Hub from file_name
id	string	Four-digit item id, e.g. 0001
question	string	Fixed question stem used for every item during evaluation
option_a	string	First answer option
option_b	string	Second answer option
answer	string	The correct option letter, A or B
answer_text	string	The text of the correct option, so the label survives any reordering
main_category	string	Interaction-pattern code: H-H, H-A, A-A, B-S
main_category_name	string	Readable form: Human-Human, Human-Animal, Animal-Animal, Bio-StillLife
subcategory	string	One of the six counter-intuitive types below
gt_prompt	string	The text-to-image prompt the scene was synthesised from

The answer letters are balanced exactly 200 / 200 across the benchmark, so a model that always answers with the same letter scores 0.50.

The prompt used in the paper

Which of the following option better describes the image?
Option A: {option_a}
Option B: {option_b}

Important Note: Answer directly with the option letter (A or B) only.

About gt_prompt

Because every scene is synthetic, the prompt that generated it is the most faithful available description of the depicted action. The paper feeds this text to a judge model instead of an image to establish an upper bound of about 0.98 accuracy, which demonstrates that the questions are logically easy once the scene is described correctly, and that failures therefore originate in perception.

Note that gt_prompt reveals the answer. It is included for reproducibility of that upper-bound experiment and should not be shown to a model under test.

Taxonomy

Interaction pattern	Counter-intuitive type	Items	Share
Human-Human (H-H)	Role Reversal - Social Power	88	22.00%
Human-Human (H-H)	Role Reversal - Kinship Care	57	14.25%
Human-Animal (H-A)	Animal Dominance over Humans	89	22.25%
Human-Animal (H-A)	Animals Providing Humanlike Care	56	14.00%
Animal-Animal (A-A)	Prey Outsmarts Predator	72	18.00%
Bio-StillLife (B-S)	Objects Act on Beings	38	9.50%
Total		400	100%

Position-bias control

The paper also evaluates a variant in which the two options are presented in the opposite order, to separate genuine understanding from a preference for a particular letter. That variant is not shipped as a separate file because it is fully reconstructible: swap option_a with option_b and flip answer.

def swap_options(example):
    example["option_a"], example["option_b"] = example["option_b"], example["option_a"]
    example["answer"] = "B" if example["answer"] == "A" else "A"
    return example

swapped = ds.map(swap_options)

Notes on the images

All 400 files are PNG. 362 are 1024x1024 and the remaining 38 are 512x512. 237 files carry an alpha channel, but it is fully opaque in every one of them, so converting to RGB loses nothing:

ds = ds.map(lambda x: {"image": x["image"].convert("RGB")})

Citation

@inproceedings{ling2026seeing,
  title     = {Seeing vs. Believing: Evaluating the Language Bias of Open-Source
               MLLMs in Counter-Intuitive Scenes},
  author    = {Ling, Chen and Zhang, Tongwei and Li, Hanqian and Ding, Nai},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural
               Language Processing (EMNLP)},
  year      = {2026}
}