pith. sign in

arxiv: 2509.07966 · v2 · submitted 2025-09-09 · 💻 cs.CV · cs.CL

Visual-TableQA: Open-Domain Benchmark for Reasoning over Table Images

Boammani Aser Lompo , Marc Haraoui This is my paper

Pith reviewed 2026-05-18 17:33 UTC · model grok-4.3

classification 💻 cs.CV cs.CL
keywords visual table reasoningmultimodal datasetsynthetic data generationtable image QAvision language model fine-tuningLLM collaborationLaTeX tablesbenchmark generalization
0
0 comments X

The pith

A new dataset of rendered table images with complex questions allows fine-tuned vision-language models to outperform several proprietary systems on external benchmarks.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Visual-TableQA as a large collection of 2,500 LaTeX-rendered table images paired with 6,000 reasoning-focused question and answer pairs. It is built through an automated pipeline where multiple language models take on different roles to generate, validate, and inspire each other in creating varied table structures and questions. Stronger models suggest initial layouts and topics, while weaker ones expand on them, and a jury of models filters for quality. When vision-language models are trained on this data, they show strong performance when tested on other table-related benchmarks, sometimes exceeding commercial models even though the training data is artificially created. This approach addresses the shortage of diverse, high-quality examples for teaching AI to understand tables in images, which appear frequently in documents and reports.

Core claim

By deploying a modular generation process that coordinates several language models through cross-model inspiration and collective filtering, the authors produce a dataset of table images and associated reasoning questions that, when used for fine-tuning, enables models to generalize effectively to independent benchmarks and surpass multiple proprietary vision-language models.

What carries the argument

The multi-model collaborative pipeline that assigns distinct roles to different language models for generation, validation, and inspiration, using stronger models to seed ideas and an LLM jury to ensure quality and diversity in the resulting table images and questions.

If this is right

  • Models fine-tuned on the dataset maintain robust performance across different table reasoning tasks outside the original data.
  • The synthetic dataset can be generated at low cost under 100 USD for the initial scale.
  • The open pipeline allows replication and expansion to create even larger collections of table reasoning examples.
  • Fine-tuned models can handle complex visual structures in tables that challenge current commercial systems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • This method of having models inspire one another could apply to generating training data for other visual reasoning domains like graphs or forms.
  • If the generated questions capture real reasoning demands, similar pipelines might reduce reliance on human-annotated datasets for multimodal tasks.
  • Further scaling could test whether such synthetic data supports training larger models specifically for document understanding.

Load-bearing premise

The data produced by the collaborative language model pipeline is diverse enough and free enough of artificial patterns to represent the kinds of tables and questions that appear in real documents and images.

What would settle it

A direct comparison showing that models fine-tuned on Visual-TableQA no longer outperform proprietary models when evaluated on a set of table images taken from actual scanned documents or web pages with typical rendering variations and noise.

Figures

Figures reproduced from arXiv: 2509.07966 by Boammani Aser Lompo, Marc Haraoui.

Figure 1
Figure 1. Figure 1: Sample question in our benchmark This second datasets category typically rely on a limited set of layout templates and involve relatively simple visual tasks or basic QA scenarios, falling short of the complexity required for thorough evaluation and advancement of reasoning capabilities. More recent efforts—such as ChartQA Masry et al., ReachQA He et al., and MATH-Vision Wang et al. (2024b)—have aimed to a… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the full pipeline architecture of Visual-TableQA. A subset of initial table images [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Correlation of model rankings on Visual-TableQA with those on three established [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Cumulative topic coverage as clusters are added by descending size. The uniform slope [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: 2D projection of the 5,000 topics using UMAP and K-Means clustering. Each color denotes [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visual diversity of table layouts. Left: seed layouts used during the first iteration of table [PITH_FULL_IMAGE:figures/full_fig_p015_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: LLM prompt used for table generation. • Text overflow: Cell content spilling outside the cell boundary, especially in narrow columns or with long strings. • Invisible content: Multirow cells with background colors that obscure cell text (e.g., white text on white background). • Improper horizontal lines: \midrule or \hline splitting across multirow cells, breaking visual coherence. Content Relevance and Co… view at source ↗
Figure 8
Figure 8. Figure 8: LLM prompt used for QA generation [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: LLM prompt used for QA evaluation. Evaluate if the answer matches the ground truth. To do so, read the question and determine whether the provided answer conveys the same meaning as the ground truth. Output a JSON response as follows: {{\"verdict\": 1}} for correct or {{\"verdict\": 0}} for incorrect [PITH_FULL_IMAGE:figures/full_fig_p018_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: LLM prompt used for Benchmark evaluation. [PITH_FULL_IMAGE:figures/full_fig_p018_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Comparison of error distributions between the pretrained [PITH_FULL_IMAGE:figures/full_fig_p020_11.png] view at source ↗
read the original abstract

Visual reasoning over structured data such as tables is a critical capability for modern vision-language models (VLMs), yet current benchmarks remain limited in scale, diversity, or reasoning depth, especially when it comes to rendered table images. Addressing this gap, we introduce Visual-TableQA, a large-scale, open-domain multimodal dataset specifically designed to evaluate and enhance visual reasoning over complex tabular data. Our generation pipeline is modular, scalable, and fully autonomous, involving multiple reasoning LLMs collaborating across distinct roles: generation, validation, and inspiration. Visual-TableQA comprises 2.5k richly structured LaTeX-rendered tables and 6k reasoning-intensive QA pairs, all produced at a cost of under USD 100. To promote diversity and creativity, our pipeline performs multi-model collaborative data generation via cross-model prompting ('inspiration') and LLM-jury filtering. Stronger models seed layouts and topics that weaker models elaborate, collectively distilling diverse reasoning patterns and visual structures into the dataset. Empirical results show that models fine-tuned on Visual-TableQA generalize robustly to external benchmarks, outperforming several proprietary models despite the dataset's synthetic nature. The full pipeline and resources are publicly available at https://github.com/AI-4-Everyone/Visual-TableQA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Visual-TableQA, an open-domain benchmark for visual reasoning over table images. It consists of 2.5k LaTeX-rendered tables and 6k QA pairs generated through a modular pipeline using multiple LLMs in collaborative roles for generation, validation, and inspiration via cross-model prompting. The central claim is that fine-tuning VLMs on this synthetic dataset leads to robust generalization on external benchmarks, with outperformance over several proprietary models.

Significance. If the empirical results hold, this provides a low-cost, scalable approach to generating diverse and reasoning-intensive table image data, addressing limitations in existing benchmarks for VLM visual reasoning on structured data. The public availability of the pipeline and dataset is a notable strength.

major comments (2)
  1. [§3] The data generation pipeline relies exclusively on clean LaTeX renders without incorporating or ablating real-world visual perturbations (e.g., perspective distortion, compression artifacts, irregular alignments). This is load-bearing for the generalization claim, as it may limit the dataset's ability to exercise the same visual parsing demands as real table images.
  2. [§4] The reported outperformance on external benchmarks lacks sufficient detail on exact metrics, baseline implementations, statistical tests, and error analysis to fully substantiate the robustness claim, particularly given the synthetic nature of the training data.
minor comments (2)
  1. [Abstract] The cost 'under USD 100' is stated but no itemized breakdown or comparison to alternative generation methods is provided in the main text.
  2. [§2] Related work section could benefit from more explicit comparison to other synthetic table datasets in terms of scale and reasoning depth.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We address each major comment below and indicate the revisions planned for the next version of the manuscript.

read point-by-point responses

  1. Referee: [§3] The data generation pipeline relies exclusively on clean LaTeX renders without incorporating or ablating real-world visual perturbations (e.g., perspective distortion, compression artifacts, irregular alignments). This is load-bearing for the generalization claim, as it may limit the dataset's ability to exercise the same visual parsing demands as real table images.

    Authors: We agree that the exclusive use of clean LaTeX renders represents a limitation for claims about robustness to real-world visual conditions. The pipeline was intentionally designed around clean renders to emphasize diversity in table structure and multi-step reasoning rather than low-level image degradation. To address the concern, we will add a new subsection in the Experiments section reporting an ablation where controlled perturbations (perspective distortion, JPEG compression, and alignment noise) are applied to a held-out portion of the test tables. We will also expand the Limitations section to explicitly discuss the scope of the generalization claims and note that future extensions will incorporate more varied real-world captures. revision: partial

  2. Referee: [§4] The reported outperformance on external benchmarks lacks sufficient detail on exact metrics, baseline implementations, statistical tests, and error analysis to fully substantiate the robustness claim, particularly given the synthetic nature of the training data.

    Authors: We acknowledge that the current presentation of results is high-level. In the revised manuscript we will expand the evaluation section to report per-benchmark accuracy and F1 scores with standard deviations across three random seeds, provide the exact prompting templates and API versions used for all proprietary baselines, include paired statistical significance tests, and add a qualitative error analysis that categorizes failure modes on the external sets. These additions will be placed in the main text or a dedicated appendix to allow readers to assess the strength of the generalization results. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; evaluation chain is independent of generation inputs.

full rationale

The paper presents an empirical dataset construction pipeline using multi-LLM collaboration to produce 2.5k LaTeX-rendered tables and 6k QA pairs, followed by fine-tuning and testing on external benchmarks. No mathematical derivations, equations, or self-referential definitions appear in the provided text. The central claim of robust generalization rests on separate external evaluation sets rather than any fitted parameter or generation output being renamed as a prediction. Self-citations are absent from the load-bearing steps, and the pipeline is described as autonomous without reducing results to its own inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claims rest on the domain assumption that current LLMs can reliably perform collaborative roles to produce high-quality diverse reasoning data and that synthetic examples transfer effectively to real-world generalization.

axioms (1)
  • domain assumption LLMs can effectively collaborate across generation, validation, and inspiration roles to produce diverse, reasoning-intensive QA pairs for rendered table images.
    The entire autonomous pipeline depends on this capability of LLMs without independent external validation of data quality beyond the reported downstream results.

pith-pipeline@v0.9.0 · 5758 in / 1382 out tokens · 53885 ms · 2026-05-18T17:33:10.951614+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TableVision: A Large-Scale Benchmark for Spatially Grounded Reasoning over Complex Hierarchical Tables

    cs.AI 2026-04 conditional novelty 7.0

    TableVision benchmark shows explicit spatial grounding recovers MLLM reasoning on hierarchical tables, delivering 12.3% accuracy improvement through a decoupled perception-reasoning framework.

Reference graph

Works this paper leans on

49 extracted references · 49 canonical work pages · cited by 1 Pith paper · 4 internal anchors

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

    work page

  2. [2]

    Magid: An automated pipeline for generating synthetic multi-modal datasets

    Hossein Aboutalebi, Hwanjun Song, Yusheng Xie, Arshit Gupta, Justin Sun, Hang Su, Igor Shalyminov, Nikolaos Pappas, Siffi Singh, and Saab Mansour. Magid: An automated pipeline for generating synthetic multi-modal datasets. arXiv preprint arXiv:2403.03194, 2024

    work page arXiv 2024

  3. [3]

    Supernova event dataset: Interpreting large language model's personality through critical event analysis

    Pranav Agarwal and Ioana Ciuc a . Supernova event dataset: Interpreting large language model's personality through critical event analysis. arXiv preprint arXiv:2506.12189, 2025

    work page arXiv 2025

  4. [4]

    Model card addendum: Claude 3.5 haiku and upgraded claude 3.5 sonnet

    Anthropic . Model card addendum: Claude 3.5 haiku and upgraded claude 3.5 sonnet. https://assets.anthropic.com/m/1cd9d098ac3e6467/original/Claude-3-Model-Card-October-Addendum.pdf, 2024. Accessed: 2025-08-01

    work page 2024

  5. [5]

    Claude opus 4 & claude sonnet 4 — system card

    Anthropic . Claude opus 4 & claude sonnet 4 — system card. https://www.anthropic.com/claude-4-system-card, May 2025. Accessed: 2025-08-01

    work page 2025

  6. [6]

    Qwen-vl: A versatile vision-language model with image, text, and box comprehension

    Shaohan Chen, Yujia Zhang, Xiangpeng Cao, Shaolei He, Chen Zhao, Zhihua Liu, Chongming Li, Jing Liu, Qiang Liu, Fan Liu, et al. Qwen-vl: A versatile vision-language model with image, text, and box comprehension. arXiv preprint arXiv:2403.18751, 2024 a . URL https://arxiv.org/abs/2403.18751

    work page arXiv 2024

  7. [7]

    Tabfact: A large-scale dataset for table-based fact verification

    Wenhu Chen, Hongmin Wang, Jianshu Chen, Yunkai Zhang, Hong Wang, Shiyang Li, Xiyou Zhou, and William Yang Wang. Tabfact: A large-scale dataset for table-based fact verification. In International Conference on Learning Representations, 2020 a . URL https://openreview.net/forum?id=rkeJRhNYDH

    work page 2020

  8. [8]

    H ybrid QA : A dataset of multi-hop question answering over tabular and textual data

    Wenhu Chen, Hanwen Zha, Zhiyu Chen, Wenhan Xiong, Hong Wang, and William Yang Wang. H ybrid QA : A dataset of multi-hop question answering over tabular and textual data. In Trevor Cohn, Yulan He, and Yang Liu (eds.), Findings of the Association for Computational Linguistics: EMNLP 2020, pp.\ 1026--1036, Online, November 2020 b . Association for Computatio...

    work page doi:10.18653/v1/2020.findings-emnlp.91 2020

  9. [9]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 24185--24198, 2024 b

    work page 2024

  10. [10]

    DeepSeek-R1-Distill-Qwen-32B

    DeepSeek-AI . DeepSeek-R1-Distill-Qwen-32B . https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B, 2025. Model card. Accessed: 2025-08-01

    work page 2025

  11. [11]

    Refocus: Visual editing as a chain of thought for structured image understanding.arXiv preprint arXiv:2501.05452,

    Xingyu Fu, Minqian Liu, Zhengyuan Yang, John Corring, Yijuan Lu, Jianwei Yang, Dan Roth, Dinei Florencio, and Cha Zhang. Refocus: Visual editing as a chain of thought for structured image understanding. arXiv preprint arXiv:2501.05452, 2025

    work page arXiv 2025

  12. [12]

    Roscoe: A suite of metrics for scoring step-by-step reasoning

    Olga Golovneva, Moya Peng Chen, Spencer Poff, Martin Corredor, Luke Zettlemoyer, Maryam Fazel-Zarandi, and Asli Celikyilmaz. Roscoe: A suite of metrics for scoring step-by-step reasoning. In The Eleventh International Conference on Learning Representations

    work page

  13. [13]

    Gemini 2.0 flash: Model card

    Google . Gemini 2.0 flash: Model card. https://storage.googleapis.com/model-cards/documents/gemini-2-flash.pdf, 2025 a . Published: 2025-04-15. Accessed: 2025-08-01

    work page 2025

  14. [14]

    Gemini 2.5 flash: Model card

    Google . Gemini 2.5 flash: Model card. https://storage.googleapis.com/model-cards/documents/gemini-2.5-flash.pdf, 2025 b . Updated: 2025-06-26. Accessed: 2025-08-01

    work page 2025

  15. [15]

    Gemini 2.5 pro: Model card

    Google . Gemini 2.5 pro: Model card. https://storage.googleapis.com/model-cards/documents/gemini-2.5-pro.pdf, 2025 c . Model card. Last updated: 2025-06-27. Accessed: 2025-08-01

    work page 2025

  16. [16]

    Distill visual chart reasoning ability from llms to mllms

    Wei He, Zhiheng Xi, Wanxu Zhao, Xiaoran Fan, Yiwen Ding, Zifei Shan, Tao Gui, Qi Zhang, and Xuanjing Huang. Distill visual chart reasoning ability from llms to mllms

    work page

  17. [17]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations

    work page

  18. [18]

    Texpert: A multi-level benchmark for evaluating latex code generation by llms

    Sahil Kale and Vijaykant Nadadur. Texpert: A multi-level benchmark for evaluating latex code generation by llms. arXiv preprint arXiv:2506.16990, 2025

    work page arXiv 2025

  19. [19]

    AIT-QA : Q uestion answering dataset over complex tables in the airline industry

    Yannis Katsis, Saneem Chemmengath, Vishwajeet Kumar, Samarth Bharadwaj, Mustafa Canim, Michael Glass, Alfio Gliozzo, Feifei Pan, Jaydeep Sen, Karthik Sankaranarayanan, and Soumen Chakrabarti. AIT-QA : Q uestion answering dataset over complex tables in the airline industry. In Anastassia Loukina, Rashmi Gangadharaiah, and Bonan Min (eds.), Proceedings of t...

    work page doi:10.18653/v1/2022.naacl-industry.34 2022

  20. [20]

    Rank correlation methods

    Maurice George Kendall. Rank correlation methods. 1948

    work page 1948

  21. [21]

    arXiv preprint arXiv:2404.19205 , year =

    Yoonsik Kim, Moonbin Yim, and Ka Yeon Song. Tablevqa-bench: A visual question answering benchmark on multiple table domains. arXiv preprint arXiv:2404.19205, 2024

    work page arXiv 2024

  22. [22]

    Thirteen ways to look at the correlation coefficient

    Joseph Lee Rodgers and W Alan Nicewander. Thirteen ways to look at the correlation coefficient. The American Statistician, 42 0 (1): 0 59--66, 1988

    work page 1988

  23. [23]

    Smir: Efficient synthetic data pipeline to improve multi-image reasoning

    Andrew Li, Rahul Thapa, Rahul Chalamala, Qingyang Wu, Kezhen Chen, and James Zou. Smir: Efficient synthetic data pipeline to improve multi-image reasoning. arXiv preprint arXiv:2501.03675, 2025

    work page arXiv 2025

  24. [24]

    LLaVA-NeXT-Interleave: Tackling Multi-image, Video, and 3D in Large Multimodal Models

    Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava-next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  25. [25]

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning

    work page

  26. [26]

    Manmatha and C

    Minesh Mathew, Dimosthenis Karatzas, R Manmatha, and CV Jawahar. Docvqa: A dataset for vqa on document images. corr abs/2007.00398 (2020). arXiv preprint arXiv:2007.00398, 2020

    work page arXiv 2007

  27. [27]

    Llama 4 Maverick 17B-128E Instruct

    Meta AI . Llama 4 Maverick 17B-128E Instruct . https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct, 2025. Model card. Accessed: 2025-08-01

    work page 2025

  28. [28]

    Mistral Small 3.1 24B Instruct

    Mistral AI . Mistral Small 3.1 24B Instruct . https://huggingface.co/mistralai/Mistral‑Small‑3.1‑24B‑Instruct‑2503, 2025. Model card. Accessed: 2025‑08‑01

    work page 2025

  29. [29]

    OpenAI. GPT-4o . https://openai.com/index/gpt-4o, 2024. Accessed: 2025-07-30

    work page 2024

  30. [30]

    OpenAI . GPT-4.1 . https://openai.com/index/gpt-4-1/, 2025 a . Accessed: 2025-08-01

    work page 2025

  31. [31]

    GPT-4o-mini

    OpenAI . GPT-4o-mini . https://platform.openai.com/docs/models/gpt-4o, 2025 b . Accessed: 2025-08-01

    work page 2025

  32. [32]

    OpenAI o3 Reasoning Model

    OpenAI. OpenAI o3 Reasoning Model . https://openai.com/index/introducing-o3-and-o4-mini/, 2025. Accessed: 2025‑07‑31

    work page 2025

  33. [33]

    doi: 10.3115/v1/P15-1142

    Panupong Pasupat and Percy Liang. Compositional semantic parsing on semi-structured tables. In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pp.\ 1470--1480, Beijing, China, July 2015. Association for Computational L...

    work page doi:10.3115/v1/p15-1142 2015

  34. [34]

    Qwen3-30B-A3B

    Qwen Team . Qwen3-30B-A3B . https://huggingface.co/Qwen/Qwen3-30B-A3B, 2025 a . Model card. Accessed: 2025-08-01

    work page 2025

  35. [35]

    Qwen3-Qwen3-32B

    Qwen Team . Qwen3-Qwen3-32B . https://huggingface.co/Qwen/Qwen3-32B, 2025 b . Model card. Accessed: 2025-08-01

    work page 2025

  36. [36]

    Qwen3-QwQ-32B

    Qwen Team . Qwen3-QwQ-32B . https://huggingface.co/Qwen/QwQ-32B, 2025 c . Model card. Accessed: 2025-08-01

    work page 2025

  37. [37]

    Reka Flash 3

    Reka AI . Reka Flash 3 . https://huggingface.co/RekaAI/reka-flash-3, 2025. Model card. Accessed: 2025-08-01

    work page 2025

  38. [38]

    Z. Z. Ren, Zhihong Shao, Junxiao Song, Huajian Xin, Haocheng Wang, Wanjia Zhao, Liyue Zhang, Zhe Fu, Qihao Zhu, Dejian Yang, Z. F. Wu, Zhibin Gou, Shirong Ma, Hongxuan Tang, Yuxuan Liu, Wenjun Gao, Daya Guo, and Chong Ruan. Deepseek-prover-v2: Advancing formal mathematical reasoning via reinforcement learning for subgoal decomposition. arXiv preprint arXi...

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2504.21801 2025

  39. [39]

    Multimodal \ qa \ : complex question answering over text, tables and images

    Alon Talmor, Ori Yoran, Amnon Catav, Dan Lahav, Yizhong Wang, Akari Asai, Gabriel Ilharco, Hannaneh Hajishirzi, and Jonathan Berant. Multimodal \ qa \ : complex question answering over text, tables and images. In International Conference on Learning Representations, 2021. URL https://openreview.net/forum?id=ee6W5UgQLa

    work page 2021

  40. [40]

    Qwen2.5-vl, January 2025

    Qwen Team. Qwen2.5-vl, January 2025. URL https://qwenlm.github.io/blog/qwen2.5-vl/

    work page 2025

  41. [41]

    Deepseek-r1t-chimera, April 2025

    TNG Technology Consulting GmbH . Deepseek-r1t-chimera, April 2025. URL https://huggingface.co/tngtech/DeepSeek-R1T-Chimera

    work page 2025

  42. [42]

    URL https://huggingface.co/datasets/cmarkea/table-vqa

    Cyrile Delestre Tom Agonnoude, 2024. URL https://huggingface.co/datasets/cmarkea/table-vqa

    work page 2024

  43. [43]

    A graph-based synthetic data pipeline for scaling high-quality reasoning instructions

    Jiankang Wang, Jianjun Xu, Xiaorui Wang, Yuxin Wang, Mengting Xing, Shancheng Fang, Zhineng Chen, Hongtao Xie, and Yongdong Zhang. A graph-based synthetic data pipeline for scaling high-quality reasoning instructions. CoRR, 2024 a

    work page 2024

  44. [44]

    Measuring multimodal mathematical reasoning with math-vision dataset

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024 b . URL https://openreview.net/forum?id=QWTCcxMpPA

    work page 2024

  45. [45]

    Grok 3 beta — the age of reasoning agents

    xAI . Grok 3 beta — the age of reasoning agents. https://x.ai/news/grok-3, 2025. Accessed: 2025-08-01

    work page 2025

  46. [46]

    MiniCPM-V: A GPT-4V Level MLLM on Your Phone

    Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  47. [47]

    Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context

    Xinyi Zheng, Douglas Burdick, Lucian Popa, Xu Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In 2021 IEEE Winter Conference on Applications of Computer Vision (WACV), pp.\ 697--706. IEEE, 2021

    work page 2021

  48. [48]

    Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning

    Victor Zhong, Caiming Xiong, and Richard Socher. Seq2sql: Generating structured queries from natural language using reinforcement learning. CoRR, abs/1709.00103, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  49. [49]

    Towards complex document understanding by discrete reasoning

    Fengbin Zhu, Wenqiang Lei, Fuli Feng, Chao Wang, Haozhou Zhang, and Tat-Seng Chua. Towards complex document understanding by discrete reasoning. In Proceedings of the 30th ACM International Conference on Multimedia, pp.\ 4857--4866, 2022

    work page 2022