cs.AISep 4, 2026

SciDocBench: A Workflow-Centered Benchmark and Data Pipeline for Scientific Document Understanding

Authors: Shenxi Wu, Yuhong Liu, Haosong Zhang, Tongjin Zou, Yanxun Zhang, Gaochang Chen, Dun Liang, Jiaqi Wang, +3 more

Organizations: The Chinese University of Hong Kong · Shanghai Artificial Intelligence Laboratory · New York University · Fudan University · Shanghai Jiao Tong University · Harbin Institute of Technology · JD Explore Academy · Centre for Perceptual and Interactive Intelligence (CPII) Limited

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Scientific papers require models to integrate evidence across text, equations, figures, tables, code, and datasets while preserving its provenance. Beyond answer correctness, scientific reading requires verifiable outputs from operations such as evidence localization, definition extraction, and consistency checking. We introduce SciDocBench, a workflow-centered benchmark targeting these operations through 124 expert-authored and difficulty-screened questions across seven research-assistant capability groups, 19 subtasks, and five scientific domains. Each question is instantiated in four matched settings formed by pairing its bilingual variants with the All Images First and Markdown Interleaved document representations, yielding 496 evaluation instances. The strongest evaluated model, Claude-Opus-5, scores 62.6 out of 100, with remaining gaps in evidence localization, structured information extraction, cross-document synthesis, and robustness to document representation. To convert these diagnostics into scalable training signals, we introduce SciDocIR, a structured representation of scientific document objects, layout and cross-reference relations, and provenance. Using SciDocIR, we construct SciDocDataset, which contains 4K supervised fine-tuning instances and 10K reinforcement-learning instances, built on 14 verifiable training subtasks. Post-training Qwen3.6-27B on task-aligned data improves its SciDocBench score from 40.03 to 45.33 with supervised fine-tuning and to 45.74 with subsequent reinforcement learning. Both adapted models preserve DocVQA and InfoVQA performance and improve ChartQA accuracy over the original model by 0.80 and 3.40 points, respectively. Together, SciDocBench, SciDocIR, and SciDocDataset connect capability diagnosis with verifiable training-data construction for scientific-document assistants.

Figures & tables

Figure 1: Overview of SciDocBench . (a) From text extraction to evidence grounding. A direct text-extraction prompt can be answered by recovering visible content, whereas the SciDocBench example additionally requires locating the document evidence supporting a scientific claim. (b) Benchmark and evaluation design. SciDocBench contains 124 expert-authored and difficulty-screened questions across five scientific domains, seven capability groups, and 19 subtasks. Its dual-axis protocol pairs bilingual question variants with the All Images First and Markdown Interleaved document representations, producing four matched settings per question. Responses are assessed using task-appropriate rule-based, LLM-as-a-judge, or execution-based evaluators. The plots summarize capability-group and evaluation-setting profiles for group-leading models, while the example illustrates the matched question and document-representation variants.
BenchmarkPrimary focusFull paper or PDFScientific objectsCross- doc.Code/data prov.
DocVQA [ 2021 ]Document-image question answering✗✗✗✗
QASPER [ 2021 ]Evidence-grounded full-paper QA✓–✗✗
ChartQA [ 2022 ]Chart question answering✗–✗✗
SCITAB [ 2023 ]Scientific-table claim verification✗✓✗✗
MMLongBench-Doc [ 2024 ]Long-context document multimodal QA✓–✗✗
DocGenome [ 2024b ]Scientific-document structure and QA✓✓✗✗
Table 1: Focused comparison with representative scientific-document benchmarks. ✓ denotes primary coverage, – partial or adjacent coverage, and ✗ that the dimension is not an evaluation target. Cross-doc. requires integrating multiple documents; code/data prov. denotes explicit reasoning over code, datasets, or provenance.
Figure 2: SciDocBench construction. Information-rich papers support expert-authored research tasks. Difficulty screening and independent verification reject trivial or ambiguous items. Bilingual and dual-representation augmentation produces four matched instances per accepted question.
Figure 3: SciDocDataset construction. Aligned PDF and LaTeX records form SciDocIR . Task-specific generation preserves locked source fields, optionally re-renders controlled evidence into page blocks, and validates instructions and targets against source or construction records.
ModelOverallEnglishChineseCapability scores
ImagesInter-ImagesInter-DocumentInformationEvidenceCross-Reconstr.Paper–Dataset
leavedleavedPerceptionExtractionVerificationDocument& Exec.CodeUnderst.
Claude-Opus-5 [ 5 ]62.6063.1961.3465.2160.6861.3755.5274.4762.8159.6551.7543.75
Claude-Opus-4.8 [ 4 ]53.4153.6352.4354.1353.4646.5957.1258.6356.6557.0944.0053.12
Claude-Sonnet-4.6 [ 6 ]49.5752.2551.1853.6241.2144.8051.5154.2547.4352.6452.1759.38
GPT-5.6-Sol [ 86 ]61.0061.0760.3261.3261.3154.0856.7878.1654.6269.7046.6959.38
Table 2: SciDocBench leaderboard. Overall is computed over all 496 evaluation instances derived from 124 underlying questions, with failed or unusable responses assigned a score of zero. Evaluation-setting scores pair English or Chinese questions with the All Images First (Images) or Markdown Interleaved (Interleaved) document representation. Capability scores correspond to the seven groups defined in Section 3 . Model families are ordered by their best overall score, and models within each family are sorted by overall score. In each score column the best, second, and third values are shaded from dark to light; ties receive the same shading.
ModelSciDocBenchGeneral document understanding
OverallCapability scoresEnglishChineseDocVQAInfoVQAChartQA
DocumentInformationEvidenceCross-Reconstr.Paper–DatasetImagesInter-ImagesInter-
PerceptionExtractionVerificationDocument& Exec.CodeUnderst.leavedleaved
Qwen3.6-27B40.0338.1438.5544.7138.4642.2840.3134.3844.6138.6135.3841.5395.8190.6774.60
+ SFT45.3340.1143.5551.4043.8552.4858.9440.6246.4842.7645.8946.2095.8091.0275.40
Improvement+5.30+1.97+5.01+6.69+5.39+10.20+18.62+6.25+1.87+4.15+10.51+4.67-0.01+0.36+0.80
Table 3: Post-training effectiveness and general-document transfer. Qwen3.6-27B and its adapted variants are evaluated on all 496 SciDocBench instances and three general document benchmarks. Input settings follow Table 2 . DocVQA and InfoVQA report validation ANLS; ChartQA reports test relaxed accuracy. Scores use final answers only; best scores are bold. Improvement rows give absolute score changes from the original model, computed before rounding; green/red indicates gains/drops.
Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

StatisticValue
Data Source
Source channelsarXiv, bioRxiv, OpenReview, PubMed
Source papers116
Benchmark Scale
Underlying questions124
Evaluation instances496
Appendix
Table 4: Benchmark and training-data statistics. Training-data totals include validation. Four views of one SFT seed remain in the same split; SFT and RL share the matched-view questions.
Figure 4: SciDocBench taxonomy and training-data distributions. A: benchmark taxonomy. B: user-prompt character counts (log scale). C: image references per instance (log-plus-one scale). Panels B–C cover 3,924 SFT and 10,143 RL instances; dots mark medians and thick bars span the interquartile range.
IDTaskEvaluation focus
A. Document Perception & Structure
A1Layout Parsing & Reading Order RecoveryRecover local reading flow in two-column PDFs with floating figures, footnotes, and cross-page continuation.
A2Element Association & Evidence LocalizationGiven a claim, locate its supporting figure, table, equation, appendix, page, or block.
A3Citation Role AnalysisIdentify whether cited work acts as background, method basis, data resource, baseline, result support, critique, or extension.
A4Dataset Lineage AnalysisRecover the source, composition, filtering, derivation, and transformation chain of datasets.
A5Figure AnalysisInterpret scientific figures, subfigures, legends, axes, annotations, and visual evidence supporting figure-level claims.
Appendix
Table 5: Full SciDocBench task taxonomy. The benchmark contains seven capability groups and nineteen subtasks.
RuleQuestionsCredit and failure conditions
Reading-order ranking5Fraction of correctly placed labels; incorrect sequence length scores zero.
Binary decisions6Fraction of correct bits in a final bitstring of the required length.
Closed scalar1Normalized label match, with the task’s answer-field/label extraction.
Grouped choices1Four groups weighted 0.30, 0.10, 0.30, 0.30; disallowed choices invalidate the affected group.
Unordered exact records1Fraction of reference dictionaries recovered exactly, ignoring key order; one-to-one matching prevents duplicate credit.
Indexed exact records1Exact row count and keys are prerequisites; schema credit plus per-row exact-match credit.
Appendix
Table 6: Closed-rule scoring beyond recursive JSON comparison. These 19 questions and the 38 recursive-JSON questions comprise the 57 R assignments.
SubtaskIDSubjectSource Document(s)Eval.
A11Computer ScienceMM-IFEngine: Towards Multimodal Instruction Following [ 24 ]J
7Nuclear TheoryRevisiting p- 11 B Fusion: Updated Cross-sections, Reactivity, and Energy Balance [ 114 ]R
13Computer ScienceDepth Anything 3: Recovering the Visual Space from Any Views [ 60 ]R
51Computer ScienceMM-IFEngine: Towards Multimodal Instruction Following [ 24 ]R
52Computer ScienceSpatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning [ 65 ]R
53MedicineQuarkMedSearch: A Long-Horizon Deep Search Agent for Exploring Medical Intelligence [ 61 ]R
Appendix
Table 7: Source documents and evaluators for Categories A1 to A2.
SubtaskIDSubjectSource Document(s)Eval.
A33Computer ScienceProximal Policy Optimization Algorithms [ 101 ]R
9BiologyProtFlow: Flow Matching-based Protein Sequence Design with Comprehensive Protein Semantic Distribution Learning and High-quality Generation [ 50 ]R
15Computer ScienceSpa3R: Predictive Spatial Field Modeling for 3D Visual Reasoning [ 44 ]R
77PhysicsQuantum computing: A taxonomy, systematic review and future directions [ 31 ]R
101PhysicsQuantum computing: A taxonomy, systematic review and future directions [ 31 ]J
103Computer ScienceQuantum computing: A taxonomy, systematic review and future directions [ 31 ]R
Appendix
Table 8: Source documents and evaluators for Categories A3 to A4.
SubtaskIDSubjectSource Document(s)Eval.
A558BiologyEstimation of the biological affinities of seven species of Sulawesi macaques based on multivariate analysis of dermatoglyphic pattern types [ 109 ]J
78ChemistryA Delocalized Mixed-Valence Dinuclear Ytterbium Complex That Displays Intervalence Charge Transfer [ 81 ]J
79Computer ScienceBiologically Plausible Learning via Bidirectional Spike-Based Distillation [ 70 ]R
80PhysicsDesign of a variable stiffness quasi-direct drive cable-actuated tensegrity robot [ 78 ]R
88BiologyDeconstruction of a spino-brain–spinal cord circuit that drives chronic pain [ 115 ]J
89PhysicsNew limits on the Pauli forbidden transitions in 12C nuclei obtained with the complete Borexino dataset [ 8 ]R
Appendix
Table 9: Source documents and evaluators for Categories A5 to B1.
SubtaskIDSubjectSource Document(s)Eval.
B25Computer ScienceSpatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning [ 65 ]R
19MedicineGlycosylation of anti-dsDNA IgG correlates with organ involvement in treatment-naive patients with systemic lupus erythematosus [ 139 ]R
20BiologyLLM-assisted systematic review of large language models in clinical medicine [ 18 ]J
21BiologyIn vivo site-specific engineering to reprogram T cells [ 80 ]J
59MathsQuantum linear system solver based on time-optimal adiabatic quantum computing and quantum approximate optimization algorithm [ 2 ]J
69ChemistryAxial chirality-induced rigidification in aminoboranes enhances persistent room-temperature phosphorescence and circularly polarized luminescence [ 28 ]J
Appendix
Table 10: Source documents and evaluators for Category B2.
SubtaskIDSubjectSource Document(s)Eval.
B36Computer ScienceProximal Policy Optimization Algorithms [ 101 ]J
22PhysicsObservation of non-Abelian topological acoustic semimetals and their phase transitions [ 43 ]J
23BiologyPoet: A generative model of protein families as sequences-of-sequences [ 112 ]J
26MathsWeighted L2L^{2} theory for the Euclidean Dirac operator in higher dimensions [ 96 ]J
70MathsWeighted L2L^{2} theory for the Euclidean Dirac operator in higher dimensions [ 96 ]J
C124Computer ScienceSpatial-SSRL: Enhancing Spatial Understanding via Self-Supervised Reinforcement Learning [ 65 ]J
Appendix
Table 11: Source documents and evaluators for Categories B3 to C2.
SubtaskIDSubjectSource Document(s)Eval.
C330MedicineGlycosylation of anti-dsDNA IgG correlates with organ involvement in treatment-naive patients with systemic lupus erythematosus [ 139 ]R
31EnvironmentSimultaneous removal of heavy metals from aqueous solutions by pineapple crown and avocado peel hydrogel composites [ 119 ]J
32MathsThe Euler Stratification for P1×P1×Pn\mathbb{P}^{1}\times\mathbb{P}^{1}\times\mathbb{P}^{n} [ 38 ]J
33MathsComputation and sampling for Schubert specializations [ 3 ]R
34MathsLearning compositional functions with transformers from easy-to-hard data [ 118 ]J
66BiologyNeural sequences underlying directed turning in Caenorhabditis elegans [ 52 ]R
Appendix
Table 12: Source documents and evaluators for Categories C3 to D1.
SubtaskIDSubjectSource Document(s)Eval.
D236Computer ScienceEagle 2.5: Boosting long-context post-training for frontier vision-language models [ 17 ] , Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale [ 35 ]J
37Computer ScienceCambrian-1: A fully open, vision-centric exploration of multimodal llms [ 110 ] , Llava-onevision: Easy visual task transfer [ 56 ] , Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale [ 35 ]R
95ChemistryStrategies for pre-training graph neural networks [ 40 ] , Geometry-enhanced molecular representation learning for property prediction [ 29 ] , Molecular contrastive learning of representations via graph neural networks [ 117 ]R
96BiologyProtFlow: Flow Matching-based Protein Sequence Design with Comprehensive Protein Semantic Distribution Learning and High-quality Generation [ 50 ] , PTM-Mamba: a PTM-aware protein language model with bidirectional gated Mamba blocks [ 89 ] , Compressing the collective knowledge of ESM into a single protein language model [ 25 ]R
109ChemistryMolecular Contrastive Learning of Representations via Graph Neural Networks [ 117 ] , ChemRL-GEM: Geometry Enhanced Molecular Representation Learning for Property Prediction [ 29 ] , Strategies for Pre-Training Graph Neural Networks [ 40 ]J
120Computer ScienceCambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs [ 110 ] , LLaVA-OneVision: Easy Visual Task Transfer [ 56 ] , MAmmoTH-VL: Eliciting Multimodal Reasoning with Instruction Tuning at Scale [ 35 ]R
Appendix
Table 13: Source documents and evaluators for Category D2.
SubtaskIDSubjectSource Document(s)Eval.
D338Computer ScienceSpatialDreamer: Incentivizing Spatial Reasoning via Active Mental Imagery [ 11 ] , Spatialladder: Progressive training for spatial reasoning in vision-language models [ 58 ]J
39Computer ScienceTs-llava: Constructing visual tokens through thumbnail-and- sampling for training-free video large language models [ 90 ] , Pllava: Parameter-free llava extension from images to videos for video dense captioning [ 130 ]J
47MedicineGlycosylation of anti-dsDNA IgG correlates with organ involvement in treatment-naive patients with systemic lupus erythematosus [ 139 ] , IgG glycans in health and disease: Prediction, intervention, prognosis, and therapy [ 104 ]R
97Computer ScienceCaprl: Stimulating dense image caption capabilities via reinforcement learning [ 126 ] , Scalecap: Inference-time scalable image captioning via dual-modality debiasing [ 127 ]R
98ChemistryHigh voltage cycling stability of LiF-coated NMC811 electrode [ 67 ] , Long-term cyclability of NCM-811 at high voltages in lithium-ion batteries: an in-depth diagnostic study [ 59 ] , Enhanced Cycling Stability of NCM811 Cathodes at High C-Rates and Voltages via LiMTFSI-Based Polymer Coating [ 48 ]R
110Computer ScienceScaleCap: Inference-Time Scalable Image Captioning via Dual-Modality Debiasing [ 127 ] , CapRL: Stimulating Dense Image Caption Capabilities via Reinforcement Learning [ 126 ]R
Appendix
Table 14: Source documents and evaluators for Category D3.
SubtaskIDSubjectSource Document(s)Eval.
E140Computer ScienceGroup Sequence Policy Optimization [ 136 ]R
41Computer ScienceLLaVA-CoT: Let Vision Language Models Reason Step-by-Step [ 129 ]J
48PhysicsStairway Codes: Floquetifying Bivariate Bicycle Codes and Beyond [ 42 ]J
49PhysicsImaginary-time evolution of interacting spin systems in the truncated Wigner approximation [ 100 ]R
68BiologyNeural sequences underlying directed turning in Caenorhabditis elegans [ 52 ]J
99BiologyCausal modelling of gene effects from regulators to programs to traits [ 88 ]R
Appendix
Table 15: Source documents and evaluators for Categories E, F, and G.
Figure 5: Expert question-authoring interface. The interface presents the source PDF and rendered pages alongside the task metadata, question, ground truth, evidence locator, and verbatim screening-model responses with author-assigned correctness labels. The populated fields constitute a representative interface record used to illustrate the annotation procedure and are not additional leaderboard results.
Figure 6: Independent-verification interface. Reviewers inspect the same PDF, task metadata, evidence locator, question, ground truth, and four screening records; audit each correctness label; complete the acceptance checklist; and record an Accept, Revise, or Reject decision.
Rater pairThree-level exactQWKBinary exactBinary κ\kappaMAE
Human expert vs. GPT-5.4-mini92.0%0.86670.0%0.3550.169
Codex vs. GPT-5.4-mini89.0%0.81880.0%0.5930.128
Human expert vs. Codex93.0%0.87274.0%0.4050.145
Appendix
Table 16: Pairwise agreement on the 100-instance semantic-judge audit. Binary results threshold the original scores at 0.5. QWK denotes quadratic-weighted Cohen’s κ\kappa , and MAE denotes mean absolute score difference.
Figure 7: Complete blind expert-review interface for the English audit item HJ067. The interface presents the item metadata, question, reference answer, anonymized model response, three-level rubric, rating controls, navigation controls, and the agreement summary revealed after all 100 ratings were completed. Model identity and the GPT-5.4-mini judgment remain hidden during rating. Browser chrome and unrelated tabs are omitted from the capture.
DirectionTraining subtasks
Layout and reading-flow recoveryBlock role classification; parent–child linking; reading-order prediction; next-hop reading-target prediction.
Table logic consistencyTable logic consistency check.
Cross-document table integrationReal cross-document table merge; single-document simulated table merge.
Chart readout and recoverySingle-point chart readout; multi-point or series readout.
Chart visual consistencyChart visual consistency.
Cross-document dataset comparisonCross-document dataset intersection.
Appendix
Table 17: Training-data construction directions and their instantiated subtasks.
SubtaskSciDocIR or document recordsExternal or synthetic sourceGPT-5.4 role and GT authority
Block role classificationPage image; target page index, bounding box, type, and layout annotationNoneLight question rewriting. The layout label remains the locked GT.
Parent–child linkingBlock text, page and bounding box; heading level; caption pairs; footnote and continuation relationsNoneQuestion rewriting and optional plausibility check. The verified relation remains GT.
Reading-order predictionOrder and reading annotations; block pages and boxes; section and discourse rolesNoneQuestion rewriting only. GT follows the stored order and layout rules.
Next-hop reading targetCurrent and successor blocks; caption pairs; figure/table references; cross-page and appendix linksNoneQuestion rewriting and optional next-hop plausibility check. The stored successor remains GT.
Table logic consistencyTable LaTeX, normalized JSON, caption, headers, cells, citing context, page, and boxNoneGenerates a natural corrupted statement, repair, and distractors, then self-checks them. Rules lock the perturbed field and answer.
Real cross-document table mergeTable schemas, rows, columns, values, captions, and nearby text from different papersNonePolishes context and question. Rules align compatible schemas and compute the merged GT.
Appendix
Table 18: Construction sources and GPT-5.4 roles for subtasks 1–7. “None” denotes no external dataset beyond the SciDocIR paper collection.
SubtaskSciDocIR or document recordsExternal or synthetic sourceGPT-5.4 role and GT authority
Single-point chart readoutFigure and caption blocks; page, box, and nearby text used for page replacementPlotQA chart, axes, series labels, and source valuesNaturalizes the question and performs repeated visual solvability checks. PlotQA values remain GT.
Multi-point or series readoutSame figure placement and context records as the single-point taskComplete PlotQA series dataRewrites the question and checks the returned list. The source series remains GT.
Chart visual consistencyFigure position, caption, and page context used to reinsert a controlled chartPlotQA values; Matplotlib radar, line, scatter, and bar renderersRewrites the question and checks solvability. Rules move one plotted value and retain the original annotation; the corruption log defines GT.
Dataset intersectionDataset-like entities mined from text, tables, and captions when available; simulated two-paper textInternal lexicon of approximately 50 CS, AI, mathematics, and natural-science resources; alias and composition poolsChecks name plausibility and polishes page prose. Exact set intersection defines GT.
Notation extractionFinal batch does not require real SciDocIR recordsMathematics, machine-learning, and physics formula templates; XeLaTeX page renderingMay polish surrounding prose while formulas and definitions stay locked. The saved symbol table defines GT.
Symbol ambiguity disambiguationFinal batch does not require real SciDocIR recordsFormula and context templates; one-column, two-column, and spanning layoutsMay naturalize local contexts without changing definitions. The rule log for each symbol occurrence defines GT.
Appendix
Table 19: Construction sources and GPT-5.4 roles for subtasks 8–14.
SettingSFTGRPO
Model adaptation and optimization
InitializationOriginal Qwen3.6-27BMerged SFT step 50
AdaptationLoRALoRA
LoRA rank / alpha32 / 6416 / 32
Learning rate1.5×10−51.5\times 10^{-5}5×10−75\times 10^{-7}
Learning-rate scheduleCosineCosine
Appendix
Table 20: Post-training configuration. The RL verifier runs on two additional H200 GPUs alongside six training GPUs.
TaskReference and predicted outputFailure interpretation
ID 13: ordered metric extraction (A1); score 0Reference: [Chamfer Distance, Precision, Recall, F1-score]. Prediction: [accuracy, completeness, Chamfer Distance (CD), precision, recall, F1-score].The model retrieves the target metrics but adds two earlier entries. Since the question restricts the section and order, every aligned position is wrong. This is a scope/ordering failure, not failure to recognize all four metrics.
ID 59: logarithmic chart readout (B2); score 0.30Reference: AQC(exp) range G ( 8 ⁣× ⁣10−48\!\times\!10^{-4} to 5 ⁣× ⁣10−15\!\times\!10^{-1} ), QAOA range C. Prediction: AQC(exp) range A ( 5 ⁣× ⁣10−35\!\times\!10^{-3} to 5 ⁣× ⁣10−15\!\times\!10^{-1} ), QAOA range C.The wrong AQC lower endpoint is also asserted in the explanation. Correct QAOA selection earns partial credit; the rubric does not credit an explanation grounded in the incorrect joint range selection.
ID 71: table consistency audit (C1); score 0.67Reference errors: 48.4, 47.8, 32.4. Prediction: 68.0, 32.4, 48.4. The missing 47.8 should be 47.3, the rounded mean of eight entries.Two errors are found, including a correctly recomputed mean, but another mean is not checked correctly. The additional highlighting complaint about 68.0 does not recover the missing reference error.
Appendix
Table 21: Selected Claude-Opus-5 failures. IDs refer to Appendix B ; outputs are abbreviated without changing the contrasted values. Scores are on the 0–1 scale.

Explore similar work

CardsList

  1. LitTraceQA: A Benchmark for Multi-Stage Grounding and Verification in Scientific Question Answering

    Aug 7, 2026Xuye Liu, Yimu Wang, Peng Shi +7Literature Review

    University of Waterloo · 2Amazon · 3City University of Hong Kong +1

  2. DocScope: Benchmarking Verifiable Reasoning for Trustworthy Long-Document Understanding

    May 9, 2026Xiang Feng, Jiawei Zhou, Zhangfeng Huang +6Multimodal DocumentsQuestion-Answering Benchmarks

    School of Computer Science, National Engineering Research Center for Multimedia Software and Hubei Key Laboratory of Multimedia and Network Communication Engineering, Wuhan University, China · Alibaba Group, Hangzhou, China · Independent Researcher +1

  3. ResearchQA: Benchmarking Citation-Grounded Question-Answering on Scientific Papers

    Jul 13, 2026Saba Imran, Debanjum Singh SolankyCitationsScholar Information Database

    Khoj Inc.