Benchmarking PDF Parsers on Table Extraction with LLM-based Semantic Evaluation
Paper • 2603.18652 • Published
pdf stringlengths 12 12 | ground_truth listlengths 14 93 |
|---|---|
pdfs/000.pdf | [
{
"type": "display-formula",
"data": "$$\\sum _{n=0}^{2^{L}-1}(\\mathbf {I} -\\mathbf {A} )^{n}=\\prod _{l=0}^{L-1}\\left(\\mathbf {I} +(\\mathbf {I} -\\mathbf {A} )^{2^{l}}\\right)$$"
},
{
"type": "text",
"data": "Pattern who what top dog player region."
},
{
"type": "inline-formula",
... |
pdfs/001.pdf | [
{
"type": "text",
"data": "Dream occur strategy level walk. Power for interest return happen follow. Mind moment term."
},
{
"type": "text",
"data": "Forget others seven heavy. Religious executive let. Sell vote thing couple crime."
},
{
"type": "text",
"data": "When now trade firm c... |
pdfs/002.pdf | [
{
"type": "display-formula",
"data": "$$x^{4}-y^{2}=0$$"
},
{
"type": "text",
"data": "Card sign order yard system see. Social almost campaign one explain social whose. Table he church response cost."
},
{
"type": "text",
"data": "State clearly weight seek not may money focus."
},
... |
pdfs/003.pdf | [
{
"type": "text",
"data": "Able fall collection certain. Pressure next science."
},
{
"type": "text",
"data": "Really become any. Worker other return line. Concern source set write as."
},
{
"type": "display-formula",
"data": "$${\\frac {1}{2}}\\omega ^{2}=-{\\frac {1}{2}}U^{2}\\left... |
pdfs/004.pdf | [
{
"type": "display-formula",
"data": "$$\\omega _{n}^{k}=e^{-{\\frac {2\\pi ik}{n}}}$$"
},
{
"type": "text",
"data": "Do always college how mean their ahead. Check song speech medical buy."
},
{
"type": "text",
"data": "Deep imagine tonight mother hear election more. Fly our on anyon... |
pdfs/005.pdf | [
{
"type": "text",
"data": "Use scientist population. House walk cell hear fast onto start."
},
{
"type": "text",
"data": "Large fly wonder strong. Night same one forward. May arrive own standard why process away. Interesting decade daughter. Consumer democratic shake professor."
},
{
"ty... |
pdfs/006.pdf | [
{
"type": "text",
"data": "Them discussion last money individual total know discuss."
},
{
"type": "inline-formula",
"data": "$0\\leq P(x_{1},x_{2},\\dots )\\leq 1$"
},
{
"type": "text",
"data": "Himself practice green require attention south compare former."
},
{
"type": "in... |
pdfs/007.pdf | [
{
"type": "text",
"data": "Somebody opportunity week make include leader yet. Media create citizen customer seat."
},
{
"type": "inline-formula",
"data": "$(1/e)v^{2}=u^{3}+(4/e-2)u^{2}+u$"
},
{
"type": "text",
"data": "Person itself health town education crime."
},
{
"type":... |
pdfs/008.pdf | [
{
"type": "display-formula",
"data": "$${\\begin{aligned}\\mathbf {f} &\\longrightarrow \\mathbf {f'} \\\\\\alpha [\\mathbf {f} ]&\\longrightarrow \\alpha [\\mathbf {f'} ]\\end{aligned}}$$"
},
{
"type": "text",
"data": "Doctor station instead among enough. Term smile participant cell best. Under... |
pdfs/009.pdf | [
{
"type": "text",
"data": "Article place writer see without keep result itself."
},
{
"type": "inline-formula",
"data": "$\\Delta s=s_{2}-s_{1}$"
},
{
"type": "text",
"data": "Anything stock light inside rich seem."
},
{
"type": "display-formula",
"data": "$$f(x)=-\\log(-... |
pdfs/010.pdf | [
{
"type": "display-formula",
"data": "$$E=g^{*}\\gamma _{\\mathbb {R} }^{n}$$"
},
{
"type": "text",
"data": "Agreement mind no husband describe watch I."
},
{
"type": "inline-formula",
"data": "$\\mu \\left(\\Omega \\cap \\mathbb {B} _{r}(p)\\right)\\leq C\\sigma \\left(\\partial \\O... |
pdfs/011.pdf | [
{
"type": "text",
"data": "Strong name court part feeling. Quickly range model community available before. Outside movement trouble floor show reality structure."
},
{
"type": "display-formula",
"data": "$$A_{\\text{new}}=A,B_{\\text{new}}=B\\oplus C,C_{\\text{new}}=C$$"
},
{
"type": "te... |
pdfs/012.pdf | [
{
"type": "display-formula",
"data": "$$\\operatorname {BTop} (\\mathbb {R} ^{n})\\to \\operatorname {BTop} (S^{n})$$"
},
{
"type": "text",
"data": "Level late let then onto. Back here hand now."
},
{
"type": "inline-formula",
"data": "$Sq^{3}=\\beta \\circ Sq^{2}\\circ r$"
},
{
... |
pdfs/013.pdf | [
{
"type": "text",
"data": "Still visit rule. Share care area nor energy."
},
{
"type": "text",
"data": "Nation even trade use score. Police building PM billion resource war light. Challenge game meet case less reality. Film any cup her."
},
{
"type": "text",
"data": "Air decade ask m... |
pdfs/014.pdf | [
{
"type": "display-formula",
"data": "$${\\text{RTC}}={\\frac {|\\Delta \\mathbf {B} _{r}|}{|\\mathbf {B} _{r}|\\Delta T}}\\times 100\\%$$"
},
{
"type": "text",
"data": "Arm true movement mouth effect. Begin control red lead happy."
},
{
"type": "inline-formula",
"data": "$\\sum _{i}... |
pdfs/015.pdf | [
{
"type": "text",
"data": "Line she possible. Prepare board than."
},
{
"type": "display-formula",
"data": "$${\\tfrac {1}{2}}({\\tfrac {1}{2}}+1)={\\tfrac {3}{4}}$$"
},
{
"type": "text",
"data": "Interview budget newspaper poor money paper perhaps. Beyond agree total plan feeling."
... |
pdfs/016.pdf | [
{
"type": "text",
"data": "Beat large miss establish reality fact raise. Level measure different me check civil. Also nice share protect quite question produce far."
},
{
"type": "text",
"data": "Source personal eight. Lose firm suggest center least young."
},
{
"type": "display-formula"... |
pdfs/017.pdf | [
{
"type": "text",
"data": "Offer several science or water question. Benefit he pressure training."
},
{
"type": "inline-formula",
"data": "$\\operatorname {pmi} (x=1;y=0)=1.584963$"
},
{
"type": "text",
"data": "I recent usually however nor capital painting."
},
{
"type": "in... |
pdfs/018.pdf | [
{
"type": "text",
"data": "Present fire under case. The guy civil eight body resource."
},
{
"type": "text",
"data": "Reason language effort change data ground."
},
{
"type": "inline-formula",
"data": "$f(r)\\equiv r^{m}\\sum _{j}g_{m,j}\\cos(j\\theta )=r^{m}\\sum _{j}g_{m,j}T_{j}(\\... |
pdfs/019.pdf | [
{
"type": "display-formula",
"data": "$$h(Z)={\\frac {1}{2}}\\log(2\\pi eN)\\,\\!$$"
},
{
"type": "text",
"data": "About door different bag. Seek tree truth."
},
{
"type": "inline-formula",
"data": "$\\phi \\mapsto P_{*}(\\phi )$"
},
{
"type": "text",
"data": "Mind series... |
pdfs/020.pdf | [
{
"type": "text",
"data": "Deal finish discover several nation understand middle."
},
{
"type": "inline-formula",
"data": "$\\lambda =(2,1)\\vdash 3$"
},
{
"type": "text",
"data": "Make third itself. Quality support point only."
},
{
"type": "inline-formula",
"data": "$\\... |
pdfs/021.pdf | [
{
"type": "display-formula",
"data": "$$p_{1}(x_{i})=f(x_{i}),p_{2}(x_{i})=(-1)^{i},i=0,...,n$$"
},
{
"type": "display-formula",
"data": "$$\\tau _{ij}=\\mu \\left({\\frac {\\partial u_{i}}{\\partial x_{j}}}+{\\frac {\\partial u_{j}}{\\partial x_{i}}}\\right)+\\delta _{ij}\\lambda {\\frac {\\par... |
pdfs/022.pdf | [
{
"type": "text",
"data": "Of good anything."
},
{
"type": "text",
"data": "Image term economic conference. Paper agree way. Describe lot physical such."
},
{
"type": "inline-formula",
"data": "$|\\Psi \\rangle ,|\\Psi '\\rangle \\in {\\mathcal {H}}_{S}\\otimes {\\mathcal {H}}_{A}$"
... |
pdfs/023.pdf | [
{
"type": "text",
"data": "Build with all live audience party."
},
{
"type": "inline-formula",
"data": "$V\\approx {\\hat {A}}_{0}\\left[f^{2}\\right]-{\\hat {A}}_{0}\\left[f\\right]^{2}$"
},
{
"type": "text",
"data": "Foreign positive cost."
},
{
"type": "inline-formula",
... |
pdfs/024.pdf | [
{
"type": "display-formula",
"data": "$$(u_{1},u_{2},e,v)$$"
},
{
"type": "text",
"data": "Thing baby item ready join. Commercial health foot make rich who."
},
{
"type": "inline-formula",
"data": "$\\sum _{i}a_{i,k}\\leq \\sum K_{i}$"
},
{
"type": "text",
"data": "Most h... |
pdfs/025.pdf | [
{
"type": "text",
"data": "Behavior smile theory yard describe majority Congress. Detail significant ahead only successful."
},
{
"type": "text",
"data": "They myself particular. Dog someone ground wife."
},
{
"type": "display-formula",
"data": "$$\\lim _{n\\to \\infty }S(n)=s3^{n}$$... |
pdfs/026.pdf | [
{
"type": "text",
"data": "Leg pick deal president official pattern."
},
{
"type": "inline-formula",
"data": "$\\mu _{B}={e\\hbar }/2m$"
},
{
"type": "text",
"data": "Course no employee several toward."
},
{
"type": "display-formula",
"data": "$$m:\\mathbb {R} ^{n}\\to \\... |
pdfs/027.pdf | [
{
"type": "text",
"data": "House at game painting avoid away story magazine."
},
{
"type": "inline-formula",
"data": "${\\text{Outcome}}(v)$"
},
{
"type": "text",
"data": "Win police speak song executive garden."
},
{
"type": "inline-formula",
"data": "$[X,Y]_{n}=[S^{n}X,... |
pdfs/028.pdf | [
{
"type": "text",
"data": "Evidence success store they administration outside."
},
{
"type": "inline-formula",
"data": "${Z_{\\mathrm {iT} }}^{2}=Z^{2}+k^{2}$"
},
{
"type": "text",
"data": "Boy make put by support."
},
{
"type": "inline-formula",
"data": "${\\textrm {vers... |
pdfs/029.pdf | [
{
"type": "text",
"data": "Arrive car either citizen style subject. Set through four seven exist enjoy experience. Question small bar will bag one per thing."
},
{
"type": "text",
"data": "His drive just six newspaper knowledge stand see. Service decision according pull. Kitchen recent most away... |
pdfs/030.pdf | [
{
"type": "text",
"data": "Suddenly choose floor beat relationship. Would benefit while hospital policy enjoy."
},
{
"type": "text",
"data": "In ask feel particularly off."
},
{
"type": "inline-formula",
"data": "$K(X)({\\sqrt[{p}]{X}})\\supset K(X)$"
},
{
"type": "text",
... |
pdfs/031.pdf | [
{
"type": "text",
"data": "Minute other spend determine rock civil thought. Fast rest bank someone shoulder explain."
},
{
"type": "text",
"data": "Happy involve treatment grow range culture off."
},
{
"type": "inline-formula",
"data": "$\\operatorname {Tails} \\left(x_{\\bullet }\\r... |
pdfs/032.pdf | [
{
"type": "display-formula",
"data": "$$\\displaystyle (ab)^{2}=a_{1}^{2}b_{2}^{2}-2a_{1}a_{2}b_{1}b_{2}+a_{2}^{2}b_{1}^{2}$$"
},
{
"type": "text",
"data": "Around where it. Couple important give choice common."
},
{
"type": "display-formula",
"data": "$$\\left({\\frac {a}{k}}\\right... |
pdfs/033.pdf | [
{
"type": "text",
"data": "Institution truth natural style protect agree think movie."
},
{
"type": "inline-formula",
"data": "$f(x_{1},...,x_{25})$"
},
{
"type": "text",
"data": "Support to his wind often between sound government."
},
{
"type": "inline-formula",
"data": ... |
pdfs/034.pdf | [
{
"type": "text",
"data": "Business offer dinner radio control difficult house if."
},
{
"type": "inline-formula",
"data": "$\\operatorname {Log} (z)=\\ln(|z|)+i\\operatorname {Arg} (z)$"
},
{
"type": "text",
"data": "Option position item wrong."
},
{
"type": "text",
"dat... |
pdfs/035.pdf | [
{
"type": "text",
"data": "Must north you crime research sign. Building piece power recently difficult political. Weight suffer out capital foot member."
},
{
"type": "text",
"data": "Laugh instead car mean fill environment outside."
},
{
"type": "inline-formula",
"data": "$K_{j-1}\\... |
pdfs/036.pdf | [
{
"type": "display-formula",
"data": "$$n_{x}<n_{y}<n_{z}$$"
},
{
"type": "text",
"data": "Leg off thank executive meeting affect."
},
{
"type": "inline-formula",
"data": "$F(p)+pF'(p)$"
},
{
"type": "text",
"data": "Itself kind enjoy great until at suddenly."
},
{
... |
pdfs/037.pdf | [
{
"type": "display-formula",
"data": "$$X_{1},\\dots ,X_{N}$$"
},
{
"type": "text",
"data": "Until establish democratic as organization individual open. Dog phone care. Source two body performance last."
},
{
"type": "text",
"data": "Nice opportunity court meet head after. Pm page en... |
pdfs/038.pdf | [
{
"type": "text",
"data": "Pattern within protect hot. Cost age follow involve between son. People rate study stop soon. The prepare beautiful get."
},
{
"type": "text",
"data": "Actually could court senior see. Turn view subject key method watch story."
},
{
"type": "inline-formula",
... |
pdfs/039.pdf | [
{
"type": "text",
"data": "Left international which evidence. Firm single beautiful least."
},
{
"type": "display-formula",
"data": "$${\\textstyle s_{n}^{2}={\\frac {1}{n-1}}\\sum _{i=1}^{n}\\left(x_{i}-{\\overline {x}}_{n}\\right)^{2}}$$"
},
{
"type": "display-formula",
"data": "$$... |
pdfs/040.pdf | [
{
"type": "text",
"data": "News its up into. Get week guess. Garden lead child full. Teach surface body court. Smile you without ask compare know believe music."
},
{
"type": "text",
"data": "South determine recent knowledge would guess analysis trouble."
},
{
"type": "inline-formula",
... |
pdfs/041.pdf | [
{
"type": "text",
"data": "Food father general TV become capital close. Hit key so serious development play organization. Public little address unit those. Particularly significant senior say population."
},
{
"type": "display-formula",
"data": "$$ds^{2}=-\\left(1-{\\frac {2Mr}{\\rho ^{2}}}\\rig... |
pdfs/042.pdf | [
{
"type": "text",
"data": "Film attack treat spend."
},
{
"type": "inline-formula",
"data": "$\\lim _{t\\to \\infty }f(t)=L$"
},
{
"type": "text",
"data": "Building they skill such theory today let. Evening role treatment than."
},
{
"type": "text",
"data": "Possible oper... |
End of preview. Expand in Data Studio
PDF Parse Bench
Benchmark for evaluating how effectively PDF parsing solutions extract mathematical formulas and tables from documents.
We generate synthetic PDFs with diverse formatting scenarios, parse them with different parsers, and score the extracted content using LLM-as-a-Judge. This semantic evaluation approach substantially outperforms traditional metrics in agreement with human judgment.
Leaderboard (2026-Q1)
Results are based on two benchmark datasets, each containing 100 synthetic PDFs:
| Parser | Tables | Formulas |
|---|---|---|
| Gemini 3 Flash | 9.50 | 9.79 |
| LightOnOCR-2-1B | 9.08 | 9.57 |
| Mistral OCR | 8.89 | 9.48 |
| dots.ocr | 8.73 | 9.55 |
| Mathpix | 8.53 | 9.66 |
| Chandra | 8.43 | 9.45 |
| Qwen3-VL-235B | 8.43 | 9.84 |
| MonkeyOCR-pro-3B | 8.39 | 9.50 |
| GLM-4.5V | 7.98 | 9.37 |
| GPT-5 mini | 7.14 | 5.57 |
| Claude Sonnet 4.6 | 7.02 | 8.50 |
| Nanonets-OCR-s | 6.92 | 9.21 |
| PP-StructureV3 | 6.86 | 9.59 |
| Gemini 2.5 Flash | 6.85 | 6.51 |
| MinerU2.5 | 6.49 | 9.32 |
| GPT-5 nano | 6.48 | 4.78 |
| DeepSeek-OCR | 5.75 | 8.97 |
| PaddleOCR-VL | 5.39 | 8.47 |
| PyMuPDF4LLM | 5.25 | 4.53 |
| GOT-OCR2.0 | 5.13 | 8.01 |
| olmOCR-2-7B | 4.05 | 9.35 |
| GROBID | 2.10 | 7.01 |
All scores are LLM-as-a-Judge ratings on a 0–10 scale, judged by Gemini 3 Flash via OpenRouter.
Datasets
2026-q1-tables-only— 100 PDFs with 451 tables (simple, moderate, complex)2026-q1-formulas-only— 100 PDFs with 1413 inline + 657 display-mode mathematical formulas
PDFs are generated synthetically using LaTeX with randomized parameters (document class, fonts, margins, column layout, line spacing). Since PDFs are generated from LaTeX source, ground truth is obtained automatically.
How to Evaluate Your Parser
pip install pdf-parse-bench
See the full evaluation guide at github.com/phorn1/pdf-parse-bench.
Why LLM-as-a-Judge?
Rule-based metrics correlate poorly with human judgment. We validated this in two human annotation studies:
- formula-metric-study — 750 human ratings: text metrics r = 0.01, CDM r = 0.31, LLM judges r = 0.74–0.82
- table-metric-study — 1,500+ human ratings: rule-based (TEDS, GriTS) top at r = 0.70, LLM judges r = 0.94
Citation
@misc{horn2025formulabench,
title = {Benchmarking Document Parsers on Mathematical Formula Extraction from PDFs},
author = {Horn, Pius and Keuper, Janis},
year = {2025},
eprint = {2511.10390},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2512.09874}
}
@misc{horn2026tablebench,
title = {Benchmarking PDF Parsers on Table Extraction with LLM-based Semantic Evaluation},
author = {Horn, Pius and Keuper, Janis},
year = {2026},
eprint = {2603.18652},
archivePrefix = {arXiv},
primaryClass = {cs.CV},
url = {https://arxiv.org/abs/2603.18652}
}
Acknowledgments
This work has been supported by the German Federal Ministry of Research, Technology and Space (BMFTR) in the program "Forschung an Fachhochschulen in Kooperation mit Unternehmen (FH-Kooperativ)" within the joint project LLMpraxis under grant 13FH622KX2.
- Downloads last month
- 13
Size of downloaded dataset files:
918 kB
Size of the auto-converted Parquet files:
823 kB
Number of rows:
200