pith. sign in

arxiv: 2601.13606 · v2 · submitted 2026-01-20 · 💻 cs.CV

ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch

Zheng Liu , Honglin Lin , Chonghan Qin , Xiaoyang Wang , Xin Gao , Yu Li , Mengzhang Cai , Yun Zhu
show 8 more authors
Zhanping Zhong Qizhi Pei Zhuoshi Pan Xiaoran Shang Bin Cui Conghui He Wentao Zhang Lijun Wu
This is my paper

Pith reviewed 2026-05-16 13:08 UTC · model grok-4.3

classification 💻 cs.CV
keywords chart reasoningvision language modelsdata synthesissynthetic datasetschain of thought reasoningprogrammatic generationcomplexity metric
0
0 comments X

The pith

A programmatic synthesis framework creates reliable complex chart data that trains an 8B vision-language model to surpass its 30B teacher.

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

The paper introduces ChartVerse to solve the shortage of good training data for chart reasoning in vision-language models. Existing synthetic charts are too simple and their question-answer pairs often contain errors or shallow reasoning. By measuring chart complexity with a new metric called Rollout Posterior Entropy and generating charts through executable code, then creating questions anchored to verifiable answers extracted from that code, the method produces large datasets of high-quality examples. Training on these datasets lets a small 8B model outperform the larger model used to generate the data.

Core claim

ChartVerse synthesizes diverse high-complexity charts using an RPE-guided chart coder that writes executable programs, then applies truth-anchored inverse QA synthesis to generate questions from deterministic answers pulled directly from the code, followed by filtering on model fail-rate and distillation of chain-of-thought reasoning. This produces ChartVerse-SFT-600K and ChartVerse-RL-40K datasets. The resulting ChartVerse-8B model achieves state-of-the-art chart reasoning performance, surpassing the Qwen3-VL-30B-A3B-Thinking teacher model and rivaling the stronger Qwen3-VL-32B-Thinking.

What carries the argument

Rollout Posterior Entropy (RPE) guided complexity-aware chart coder combined with truth-anchored inverse QA synthesis, which extracts answers from source code before generating questions to ensure consistency.

If this is right

  • Produces scalable datasets of 600K supervised fine-tuning samples and 40K reinforcement learning samples.
  • Trains an 8B model to exceed the performance of its 30B teacher on chart reasoning benchmarks.
  • Generalizes the approach to create reliable reasoning data without human annotation.
  • Releases code, weights, and datasets for further use in vision-language model training.

Where Pith is reading between the lines

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

  • Similar synthesis pipelines could address data scarcity in other visual reasoning tasks such as diagram or table interpretation.
  • Models might eventually use this method to generate their own training data in a self-improving loop.
  • The emphasis on code-executable charts suggests potential for automatic verification in broader multimodal tasks.

Load-bearing premise

The synthesized charts and reasoning chains capture the essential complexities and logical structures found in real-world chart reasoning problems.

What would settle it

If a new test set of human-designed complex charts shows the 8B model underperforming the teacher model, that would indicate the synthetic data does not fully transfer.

Figures

Figures reproduced from arXiv: 2601.13606 by Bin Cui, Chonghan Qin, Conghui He, Honglin Lin, Lijun Wu, Mengzhang Cai, Qizhi Pei, Wentao Zhang, Xiaoran Shang, Xiaoyang Wang, Xin Gao, Yu Li, Yun Zhu, Zhanping Zhong, Zheng Liu, Zhuoshi Pan.

Figure 1
Figure 1. Figure 1: Illustration of Rollout Posterior Entropy. We quantify complexity via generative stability: [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of ChartVerse. Stage 1 trains a complexity-aware chart coder via RPE-guided [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance evolution of ChartVerse across different training phases on chart and STEM [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation study results on different datasets and the proposed RPE strategy. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: t-SNE visualizations of feature distributions for ChartVerse-SFT and existing datasets. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the ChartVerse-SFT dataset. The samples demonstrate high diversity in chart [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: ChartVerse-SFT QA Example-1 21 [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: ChartVerse-SFT QA Example-2 22 [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: ChartVerse-SFT QA Example-3 23 [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
read the original abstract

Chart reasoning is a critical capability for Vision Language Models (VLMs). However, the development of open-source models is severely hindered by the lack of high-quality training data. Existing datasets suffer from a dual challenge: synthetic charts are often simplistic and repetitive, while the associated QA pairs are prone to hallucinations and lack the reasoning depth required for complex tasks. To bridge this gap, we propose ChartVerse, a scalable framework designed to synthesize complex charts and reliable reasoning data from scratch. (1) To address the bottleneck of simple patterns, we first introduce Rollout Posterior Entropy (RPE), a novel metric that quantifies chart complexity. Guided by RPE, we develop complexity-aware chart coder to autonomously synthesize diverse, high-complexity charts via executable programs. (2) To guarantee reasoning rigor, we develop truth-anchored inverse QA synthesis. Diverging from standard generation, we adopt an answer-first paradigm: we extract deterministic answers directly from the source code, generate questions conditional on these anchors, and enforce strict consistency verification. To further elevate difficulty and reasoning depth, we filter samples based on model fail-rate and distill high-quality Chain-of-Thought (CoT) reasoning. We curate ChartVerse-SFT-600K and ChartVerse-RL-40K using Qwen3-VL-30B-A3B-Thinking as the teacher. Experimental results demonstrate that ChartVerse-8B achieves state-of-the-art performance, notably surpassing its teacher and rivaling the stronger Qwen3-VL-32B-Thinking. We release our code, model weights, and datasets in https://chartverse.github.io.

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

3 major / 2 minor

Summary. The paper introduces ChartVerse, a framework for scalable synthesis of complex charts and reliable reasoning data from scratch to address data scarcity in VLM chart reasoning. It proposes Rollout Posterior Entropy (RPE) to guide complexity-aware programmatic chart generation, truth-anchored inverse QA synthesis (answer-first extraction from code with consistency verification), model fail-rate filtering, and CoT distillation from Qwen3-VL-30B-A3B-Thinking. This yields ChartVerse-SFT-600K and ChartVerse-RL-40K datasets; the resulting ChartVerse-8B model reportedly achieves SOTA performance, surpassing the teacher and rivaling Qwen3-VL-32B-Thinking.

Significance. If the performance claims hold under rigorous verification, the work would be significant for the field: it offers a programmatic route to high-complexity, hallucination-resistant chart reasoning data that scales beyond existing simplistic synthetic sets, directly supporting open-source VLM progress on visual reasoning tasks. The public release of code, weights, and datasets strengthens reproducibility and downstream utility.

major comments (3)
  1. [Abstract] Abstract: The headline claim that ChartVerse-8B surpasses its teacher Qwen3-VL-30B-A3B-Thinking and rivals Qwen3-VL-32B-Thinking is stated without any numerical metrics, error rates, ablation tables, or verification statistics on the consistency checks; this absence makes the central empirical result impossible to assess from the given text.
  2. [§3] §3 (RPE-guided synthesis and filtering): The RPE metric is presented as the key innovation for selecting high-complexity charts, yet no quantitative correlation is shown between RPE scores and downstream reasoning depth, nor any ablation comparing RPE-guided vs. random or heuristic filtering on held-out chart distributions.
  3. [Experiments] Experimental section: The pipeline uses the same teacher both to filter hard examples and to distill CoT, but no ablations are reported on real-world (non-synthetic) charts, alternative teacher models, or distribution-shift tests; this leaves the transferability assumption untested and load-bearing for the generalization claim.
minor comments (2)
  1. [§3.1] The formal definition and rollout procedure for RPE would benefit from an explicit equation or algorithm box to support independent re-implementation.
  2. [Figures/Tables] Table captions and axis labels in the results figures should explicitly state the evaluation benchmarks and whether numbers are reported on synthetic or real charts.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below and will incorporate revisions to improve clarity and empirical support.

read point-by-point responses

  1. Referee: [Abstract] The headline claim that ChartVerse-8B surpasses its teacher Qwen3-VL-30B-A3B-Thinking and rivals Qwen3-VL-32B-Thinking is stated without any numerical metrics, error rates, ablation tables, or verification statistics on the consistency checks.

    Authors: We agree that the abstract should be self-contained with key numbers. In revision we will add specific metrics (e.g., absolute accuracy gains on ChartQA, PlotQA, and DocVQA), the pass rate of consistency verification (>92%), and a one-sentence summary of the main ablation outcomes. revision: yes

  2. Referee: [§3] The RPE metric is presented as the key innovation yet no quantitative correlation is shown between RPE scores and downstream reasoning depth, nor any ablation comparing RPE-guided vs. random or heuristic filtering on held-out chart distributions.

    Authors: We will add a correlation analysis (RPE vs. teacher fail-rate and reasoning-step count) and a new ablation table in §3.3 that compares RPE-guided synthesis against random sampling and heuristic baselines on held-out chart distributions, reporting downstream accuracy deltas. revision: yes

  3. Referee: [Experiments] The pipeline uses the same teacher both to filter hard examples and to distill CoT, but no ablations are reported on real-world (non-synthetic) charts, alternative teacher models, or distribution-shift tests.

    Authors: We will expand the experimental section with direct evaluations on real-world chart benchmarks (ChartQA, PlotQA) to demonstrate transfer. Alternative-teacher and explicit distribution-shift ablations will be added where computationally feasible; remaining cases will be discussed as limitations with supporting evidence from the current teacher-controlled setting. revision: partial

Circularity Check

0 steps flagged

No significant circularity detected in claimed derivation

full rationale

The paper presents an empirical pipeline for data synthesis (RPE-guided chart generation, answer-first QA extraction from code, fail-rate filtering, and teacher CoT distillation) followed by training and benchmarking of ChartVerse-8B. No equations, definitions, or steps reduce the reported performance numbers to the synthetic inputs by construction; the final SOTA claim is an external measurement against independent test distributions rather than a tautological restatement of the teacher outputs or fitted parameters. The use of the same teacher for data curation is a standard distillation choice and does not create a self-definitional or load-bearing self-citation loop within the derivation itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the assumption that executable chart code provides ground-truth answers and that RPE correlates with reasoning difficulty; no free parameters are explicitly fitted in the abstract description.

axioms (1)
  • domain assumption Executable chart generation code yields deterministic, hallucination-free answers
    Invoked when the paper extracts answers directly from source code before generating questions.
invented entities (1)
  • Rollout Posterior Entropy (RPE) no independent evidence
    purpose: Quantify chart complexity to guide synthesis toward diverse, high-difficulty samples
    New metric introduced to select charts; no independent evidence provided beyond its use in the pipeline.

pith-pipeline@v0.9.0 · 5641 in / 1312 out tokens · 25099 ms · 2026-05-16T13:08:45.625790+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 2 Pith papers

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

  1. ConeSep: Cone-based Robust Noise-Unlearning Compositional Network for Composed Image Retrieval

    cs.CV 2026-04 unverdicted novelty 7.0

    ConeSep tackles noisy triplet correspondences in composed image retrieval by introducing geometric fidelity quantization to locate noise, negative boundary learning for semantic opposites, and targeted unlearning via ...

  2. Air-Know: Arbiter-Calibrated Knowledge-Internalizing Robust Network for Composed Image Retrieval

    cs.CV 2026-04 unverdicted novelty 6.0

    Air-Know decouples MLLM-based external arbitration from proxy learning via knowledge internalization and dual-stream training to overcome noisy triplet correspondence in composed image retrieval.

Reference graph

Works this paper leans on

68 extracted references · 68 canonical work pages · cited by 2 Pith papers · 13 internal anchors

  1. [1]

    Claude 3.5 technical report

    Anthropic. Claude 3.5 technical report. Technical report, Anthropic, 2025. URL https://www.anthropic. com/news/claude-3-family

    work page 2025

  2. [2]

    Qwen3-vl technical report,

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    work page

  3. [3]

    URLhttps://arxiv.org/abs/2511.21631

    work page internal anchor Pith review Pith/arXiv arXiv

  4. [4]

    Opendataarena: A fair and open arena for benchmarking post-training dataset value, 2025

    Mengzhang Cai, Xin Gao, Yu Li, Honglin Lin, Zheng Liu, Zhuoshi Pan, Qizhi Pei, Xiaoran Shang, Mengyuan Sun, Zinan Tang, Xiaoyang Wang, Zhanping Zhong, Yun Zhu, Dahua Lin, Conghui He, and Lijun Wu. Opendataarena: A fair and open arena for benchmarking post-training dataset value, 2025. URL https: //arxiv.org/abs/2512.14051

    work page arXiv 2025

  5. [5]

    Chart-r1: Chain-of- thought supervision and reinforcement for advanced chart reasoner.arXiv preprint arXiv:2507.15509, 2025

    Lei Chen, Xuanle Zhao, Zhixiong Zeng, Jing Huang, Yufeng Zhong, and Lin Ma. Chart-r1: Chain-of-thought supervision and reinforcement for advanced chart reasoner, 2025. URL https://arxiv.org/abs/2507.15509

    work page arXiv 2025

  6. [6]

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models, 2024

    Haodong Duan, Xinyu Fang, Junming Yang, Xiangyu Zhao, Yuxuan Qiao, Mo Li, Amit Agarwal, Zhe Chen, Lin Chen, Yuan Liu, Yubo Ma, Hailong Sun, Yifan Zhang, Shiyin Lu, Tack Hwa Wong, Weiyun Wang, Peiheng Zhou, Xiaozhe Li, Chaoyou Fu, Junbo Cui, Jixuan Chen, Enxin Song, Song Mao, Shengyuan Ding, Tianhao Liang, Zicheng Zhang, Xiaoyi Dong, Yuhang Zang, Pan Zhang...

    work page arXiv 2025

  7. [7]

    Gemini 3 technical report

    Google DeepMind. Gemini 3 technical report. Technical report, Google, 2025. URL https://storage. googleapis.com/deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf

    work page 2025

  8. [8]

    Chartllama: A mul- timodal llm for chart understanding and generation.arXiv preprint arXiv:2311.16483, 2023

    Yucheng Han, Chi Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartllama: A multimodal llm for chart understanding and generation, 2023. URL https://arxiv.org/abs/ 2311.16483

    work page arXiv 2023

  9. [9]

    Distill visual chart reasoning ability from llms to mllms, 2025

    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, 2025. URL https://arxiv.org/abs/2410. 18798

    work page 2025

  10. [10]

    Evochart: A benchmark and a self-training approach towards real-world chart understanding, 2025

    Muye Huang, Han Lai, Xinyu Zhang, Wenjun Wu, Jie Ma, Lingling Zhang, and Jun Liu. Evochart: A benchmark and a self-training approach towards real-world chart understanding, 2025. URL https://arxiv. org/abs/2409.01577

    work page arXiv 2025

  11. [11]

    Qwen2.5-Coder Technical Report

    Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, Kai Dang, Yang Fan, Yichang Zhang, An Yang, Rui Men, Fei Huang, Bo Zheng, Yibo Miao, Shanghaoran Quan, Yunlong Feng, Xingzhang Ren, Xuancheng Ren, Jingren Zhou, and Junyang Lin. Qwen2.5-coder technical report, 2024. URLhttps://arxiv.org/ab...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  12. [12]

    Chartreasoner: Code-driven modality bridging for long-chain reasoning in chart question answering, 2025

    Caijun Jia, Nan Xu, Jingxuan Wei, Qingli Wang, Lei Wang, Bihui Yu, and Junnan Zhu. Chartreasoner: Code-driven modality bridging for long-chain reasoning in chart question answering, 2025. URL https: //arxiv.org/abs/2506.10116

    work page arXiv 2025

  13. [13]

    Chart-coca: Self-improving chart understanding of vision lms via code-driven synthesis and candidate-conditioned answering, 2025

    Gongyao Jiang and Qiong Luo. Chart-coca: Self-improving chart understanding of vision lms via code-driven synthesis and candidate-conditioned answering, 2025. URLhttps://arxiv.org/abs/2508.11975

    work page arXiv 2025

  14. [14]

    DVQA: Understanding Data Visualizations via Question Answering

    Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering, 2018. URLhttps://arxiv.org/abs/1801.08163

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  15. [15]

    FigureQA: An Annotated Figure Dataset for Visual Reasoning

    Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning, 2018. URL https://arxiv.org/abs/ 1710.07300. 12 ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  16. [16]

    Chartgen: Scaling chart understanding via code-guided synthetic chart generation, 2025

    Jovana Kondic, Pengyuan Li, Dhiraj Joshi, Zexue He, Shafiq Abedin, Jennifer Sun, Ben Wiesel, Eli Schwartz, Ahmed Nassar, Bo Wu, Assaf Arbelle, Aude Oliva, Dan Gutfreund, Leonid Karlinsky, and Rogerio Feris. Chartgen: Scaling chart understanding via code-guided synthetic chart generation, 2025. URL https: //arxiv.org/abs/2507.19492

    work page arXiv 2025

  17. [17]

    Efficient Memory Management for Large Language Model Serving with PagedAttention

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gon- zalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention, 2023. URLhttps://arxiv.org/abs/2309.06180

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  18. [18]

    Mmc: Advancing multimodal chart understand- ing with large-scale instruction tuning

    Fuxiao Liu, Xiaoyang Wang, Wenlin Yao, Jianshu Chen, Kaiqiang Song, Sangwoo Cho, Yaser Yacoob, and Dong Yu. Mmc: Advancing multimodal chart understanding with large-scale instruction tuning, 2024. URL https://arxiv.org/abs/2311.10774

    work page arXiv 2024

  19. [19]

    Wong, Songyang Zhang, and Kai Chen

    Shudong Liu, Hongwei Liu, Junnan Liu, Linchen Xiao, Songyang Gao, Chengqi Lyu, Yuzhe Gu, Wenwei Zhang, Derek F. Wong, Songyang Zhang, and Kai Chen. Compassverifier: A unified and robust verifier for llms evaluation and outcome reward, 2025. URLhttps://arxiv.org/abs/2508.03686

    work page arXiv 2025

  20. [20]

    Synthvlm: High-efficiency and high-quality syn- thetic data for vision language models

    Zheng Liu, Hao Liang, Bozhou Li, Wentao Xiong, Chong Chen, Conghui He, Wentao Zhang, and Bin Cui. Synthvlm: Towards high-quality and efficient synthesis of image-caption datasets for vision-language models, 2025. URLhttps://arxiv.org/abs/2407.20756

    work page arXiv 2025

  21. [21]

    FLARE: Fully Integration of Vision-Language Representations for Deep Cross-Modal Understanding

    Zheng Liu, Mengjie Liu, Jingzhou Chen, Jingwei Xu, Bin Cui, Conghui He, and Wentao Zhang. Fusion: Fully integration of vision-language representations for deep cross-modal understanding, 2025. URL https://arxiv.org/abs/2504.09925

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  22. [22]

    Start: Spatial and textual learning for chart understanding, 2025

    Zhuoming Liu, Xiaofeng Gao, Feiyang Niu, Qiaozi Gao, Liu Liu, and Robinson Piramuthu. Start: Spatial and textual learning for chart understanding, 2025. URLhttps://arxiv.org/abs/2512.07186

    work page arXiv 2025

  23. [23]

    MathVista: Evaluating Mathematical Reasoning of Foundation Models in Visual Contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts.arXiv preprint arXiv:2310.02255, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  24. [24]

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning, 2022

    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, 2022. URL https://arxiv.org/abs/2203. 10244

    work page 2022

  25. [25]

    Chartinstruct: Instruction tuning for chart comprehension and reasoning

    Ahmed Masry, Mehrad Shahmohammadi, Md Rizwan Parvez, Enamul Hoque, and Shafiq Joty. Chartinstruct: Instruction tuning for chart comprehension and reasoning, 2024. URL https://arxiv.org/abs/2403.09028

    work page arXiv 2024

  26. [26]

    Chart- gemma: Visual instruction-tuning for chart reasoning in the wild, 2024

    Ahmed Masry, Megh Thakkar, Aayush Bajaj, Aaryaman Kartha, Enamul Hoque, and Shafiq Joty. Chart- gemma: Visual instruction-tuning for chart reasoning in the wild, 2024. URL https://arxiv.org/abs/2407. 04172

    work page 2024

  27. [27]

    Chartqapro: A more di- verse and challenging benchmark for chart question answer- ing.arXiv preprint arXiv:2504.05506, 2025

    Ahmed Masry, Mohammed Saidul Islam, Mahir Ahmed, Aayush Bajaj, Firoz Kabir, Aaryaman Kartha, Md Tahmid Rahman Laskar, Mizanur Rahman, Shadikur Rahman, Mehrad Shahmohammadi, Megh Thakkar, Md Rizwan Parvez, Enamul Hoque, and Shafiq Joty. Chartqapro: A more diverse and challenging benchmark for chart question answering, 2025. URLhttps://arxiv.org/abs/2504.05506

    work page arXiv 2025

  28. [28]

    Chartassisstant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning,

    Fanqing Meng, Wenqi Shao, Quanfeng Lu, Peng Gao, Kaipeng Zhang, Yu Qiao, and Ping Luo. Chartassisstant: A universal chart multimodal language model via chart-to-table pre-training and multitask instruction tuning,

    work page

  29. [29]

    URLhttps://arxiv.org/abs/2401.02384

    work page arXiv

  30. [30]

    Khapra, and Pratyush Kumar

    Nitesh Methani, Pritha Ganguly, Mitesh M. Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots, 2020. URLhttps://arxiv.org/abs/1909.00997

    work page arXiv 2020

  31. [31]

    MinerU2.5: A Decoupled Vision-Language Model for Efficient High-Resolution Document Parsing

    Junbo Niu, Zheng Liu, Zhuangcheng Gu, Bin Wang, Linke Ouyang, Zhiyuan Zhao, Tao Chu, Tianyao He, Fan Wu, Qintong Zhang, Zhenjiang Jin, Guang Liang, Rui Zhang, Wenzheng Zhang, Yuan Qu, Zhifei Ren, Yuefeng Sun, Yuanhong Zheng, Dongsheng Ma, Zirui Tang, Boyu Niu, Ziyang Miao, Hejun Dong, Siyi Qian, Junyuan Zhang, Jingzhou Chen, Fangdong Wang, Xiaomeng Zhao, ...

    work page internal anchor Pith review arXiv 2025

  32. [32]

    Gpt-5 system card

    OpenAI. Gpt-5 system card. Technical report, OpenAI, 2025. URL https://cdn.openai.com/ gpt-5-system-card.pdf

    work page 2025

  33. [33]

    Learning Transferable Visual Models From Natural Language Supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. URLhttps://arxiv.org/abs/2103.00020

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  34. [34]

    Hybridflow: A flexible and efficient rlhf framework

    Guangming Sheng, Chi Zhang, Zilingfeng Ye, Xibin Wu, Wang Zhang, Ru Zhang, Yanghua Peng, Haibin Lin, and Chuan Wu. Hybridflow: A flexible and efficient rlhf framework. InProceedings of the Twentieth European Conference on Computer Systems, EuroSys ’25, page 1279–1297. ACM, March 2025. doi: 10.1145/ 3689031.3696075. URLhttp://dx.doi.org/10.1145/3689031.3696075

    work page doi:10.1145/3689031.3696075 2025

  35. [35]

    Chartmuseum: Testing visual reasoning capabilities of large vision-language models, 2025

    Liyan Tang, Grace Kim, Xinyu Zhao, Thom Lake, Wenxuan Ding, Fangcong Yin, Prasann Singhal, Manya Wadhwa, Zeyu Leo Liu, Zayne Sprague, Ramya Namuduri, Bodun Hu, Juan Diego Rodriguez, Puyuan Peng, and Greg Durrett. Chartmuseum: Testing visual reasoning capabilities of large vision-language models, 2025. URLhttps://arxiv.org/abs/2505.13444

    work page arXiv 2025

  36. [36]

    Internvl3.5: Advancing open-source multimodal models in versatility, reasoning, and efficiency,

    Weiyun Wang, Zhangwei Gao, Lixin Gu, Hengjun Pu, Long Cui, Xingguang Wei, Zhaoyang Liu, Linglin Jing, Shenglong Ye, Jie Shao, Zhaokai Wang, Zhe Chen, Hongjie Zhang, Ganlin Yang, Haomin Wang, Qi Wei, Jinhui Yin, Wenhao Li, Erfei Cui, Guanzhou Chen, Zichen Ding, Changyao Tian, Zhenyu Wu, Jingjing Xie, Zehao Li, Bowen Yang, Yuchen Duan, Xuehui Wang, Zhi Hou,...

    work page

  37. [37]

    URLhttps://arxiv.org/abs/2508.18265

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    Charxiv: Charting gaps in realistic chart understanding in multimodal llms.arXiv preprint arXiv:2406.18521, 2024

    Zirui Wang, Mengzhou Xia, Luxi He, Howard Chen, Yitao Liu, Richard Zhu, Kaiqu Liang, Xindi Wu, Haotian Liu, Sadhika Malladi, Alexis Chevalier, Sanjeev Arora, and Danqi Chen. Charxiv: Charting gaps in realistic chart understanding in multimodal llms, 2024. URLhttps://arxiv.org/abs/2406.18521

    work page arXiv 2024

  39. [39]

    Chartx & chartvlm: A versatile bench- mark and foundation model for complicated chart reasoning

    Renqiu Xia, Bo Zhang, Hancheng Ye, Xiangchao Yan, Qi Liu, Hongbin Zhou, Zijun Chen, Peng Ye, Min Dou, Botian Shi, Junchi Yan, and Yu Qiao. Chartx & chartvlm: A versatile benchmark and foundation model for complicated chart reasoning, 2025. URLhttps://arxiv.org/abs/2402.12185

    work page arXiv 2025

  40. [40]

    LogicVista: Multimodal LLM Logical Reasoning Benchmark in Visual Contexts

    Yijia Xiao, Edward Sun, Tianyu Liu, and Wei Wang. Logicvista: Multimodal llm logical reasoning benchmark in visual contexts.arXiv preprint arXiv:2407.04973, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  41. [41]

    Chartm3: A multi-stage code-driven pipeline for constructing multi-dimensional and multi-step visual reasoning data in chart comprehension, 2025

    Duo Xu, Hao Cheng, Xin Lin, Zhen Xie, and Hao Wang. Chartm3: A multi-stage code-driven pipeline for constructing multi-dimensional and multi-step visual reasoning data in chart comprehension, 2025. URL https://arxiv.org/abs/2511.02415

    work page arXiv 2025

  42. [42]

    Visulogic: A benchmark for evaluating visual reasoning in multi-modal large language models.arXiv preprint arXiv:2504.15279, 2025

    Weiye Xu, Jiahao Wang, Weiyun Wang, Zhe Chen, Wengang Zhou, Aijun Yang, Lewei Lu, Houqiang Li, Xiaohua Wang, Xizhou Zhu, et al. Visulogic: A benchmark for evaluating visual reasoning in multi-modal large language models.arXiv preprint arXiv:2504.15279, 2025

    work page arXiv 2025

  43. [43]

    Chartbench: A benchmark for complex visual reasoning in charts.arXiv preprint arXiv:2312.15915,

    Zhengzhuo Xu, Sinan Du, Yiyan Qi, Chengjin Xu, Chun Yuan, and Jian Guo. Chartbench: A benchmark for complex visual reasoning in charts, 2024. URLhttps://arxiv.org/abs/2312.15915

    work page arXiv 2024

  44. [44]

    Scaling text-rich image understanding via code-guided synthetic multimodal data generation, 2025

    Yue Yang, Ajay Patel, Matt Deitke, Tanmay Gupta, Luca Weihs, Andrew Head, Mark Yatskar, Chris Callison- Burch, Ranjay Krishna, Aniruddha Kembhavi, and Christopher Clark. Scaling text-rich image understanding via code-guided synthetic multimodal data generation, 2025. URLhttps://arxiv.org/abs/2502.14846

    work page arXiv 2025

  45. [45]

    Effective training data synthesis for improving mllm chart understanding, 2025

    Yuwei Yang, Zeyu Zhang, Yunzhong Hou, Zhuowan Li, Gaowen Liu, Ali Payani, Yuan-Sen Ting, and Liang Zheng. Effective training data synthesis for improving mllm chart understanding, 2025. URL https://arxiv.org/abs/2508.06492. 14 ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch

    work page arXiv 2025

  46. [46]

    Open- mmreasoner: Pushing the frontiers for multimodal rea- soning with an open and general recipe.arXiv preprint arXiv:2511.16334, 2025

    Kaichen Zhang, Keming Wu, Zuhao Yang, Kairui Hu, Bin Wang, Ziwei Liu, Xingxuan Li, and Lidong Bing. Openmmreasoner: Pushing the frontiers for multimodal reasoning with an open and general recipe.arXiv preprint arXiv:2511.16334, 2025

    work page arXiv 2025

  47. [47]

    Tinychart: Efficient chart understanding with visual token merging and program- of-thoughts learning.arXiv preprint arXiv:2404.16635,

    Liang Zhang, Anwen Hu, Haiyang Xu, Ming Yan, Yichen Xu, Qin Jin, Ji Zhang, and Fei Huang. Tinychart: Efficient chart understanding with visual token merging and program-of-thoughts learning, 2024. URL https://arxiv.org/abs/2404.16635

    work page arXiv 2024

  48. [48]

    Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? InEuropean Conference on Computer Vision, pages 169–186

    Renrui Zhang, Dongzhi Jiang, Yichi Zhang, Haokun Lin, Ziyu Guo, Pengshuo Qiu, Aojun Zhou, Pan Lu, Kai-Wei Chang, Yu Qiao, et al. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? InEuropean Conference on Computer Vision, pages 169–186. Springer, 2024

    work page 2024

  49. [49]

    Chartcoder: Advancing multimodal large language model for chart-to-code generation, 2025

    Xuanle Zhao, Xianzhen Luo, Qi Shi, Chi Chen, Shuo Wang, Zhiyuan Liu, and Maosong Sun. Chartcoder: Advancing multimodal large language model for chart-to-code generation, 2025. URL https://arxiv.org/ abs/2501.06598

    work page arXiv 2025

  50. [50]

    Group Sequence Policy Optimization

    Chujie Zheng, Shixuan Liu, Mingze Li, Xiong-Hui Chen, Bowen Yu, Chang Gao, Kai Dang, Yuqiong Liu, Rui Men, An Yang, Jingren Zhou, and Junyang Lin. Group sequence policy optimization, 2025. URL https://arxiv.org/abs/2507.18071

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  51. [51]

    LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models, 2024. URL https://arxiv.org/abs/ 2403.13372

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  52. [52]

    Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models.arXiv preprint arXiv:2411.00836, 2024

    Chengke Zou, Xingang Guo, Rui Yang, Junyu Zhang, Bin Hu, and Huan Zhang. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models.arXiv preprint arXiv:2411.00836, 2024. 15 ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch Appendix A More Training Details We implement...

    work page arXiv 2024

  53. [53]

    ## Design Constraints (Difficulty Control) •Target Level: The logic must beaccessible to a High School student

    Implement: Write a standalone Python script that executes this logic to produce a single significant result (a specific number or a text label). ## Design Constraints (Difficulty Control) •Target Level: The logic must beaccessible to a High School student. •Linearity: Focus on clear, linear logic, but avoid trivial operations. ## Logic Architecture (The ”...

    work page

  54. [54]

    3.Structural Dependency: Do not simply process the dataset uniformly

    Multi-Subplot Integration (CRITICAL): If the input code contains multiple subplots, your calcula- tion logicMUST synthesize data across at least 2 subplots. 3.Structural Dependency: Do not simply process the dataset uniformly

    work page

  55. [55]

    ## Output Requirements 1.Data Context: A brief professional summary of what the data represents

    Derived Aggregation: The final output must be a single synthesized value (numerical scalar OR categorical label). ## Output Requirements 1.Data Context: A brief professional summary of what the data represents. 2.Logic Blueprint: A step-by-step text description of the algorithm. 3.The Python Script (The Solution): •Self-Contained: Define all data lists ex...

    work page

  56. [56]

    </question>

    The Analytical Question: The abstract, scenario-based question wrapped in <question>... </question>. ## Input Data

    work page

  57. [57]

    Original Visualization Code: {chart code}

    work page

  58. [58]

    Your task is to answer a specific question by interpreting the raw data structures defined in a Python code snippet

    The Solution Script (The Answer Logic): {generated python code} 27 ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch Table 12: Prompt for Inverse Logic Construction - Step 3: LLM Inference Step 3: LLM Inference Prompt ## Role You are an expert inChart Code ComprehensionandData Reasoning. Your task is to answer a specific...

    work page

  59. [59]

    You do not need to ”run” the code, but you must ”read” and understand the variables (lists, dictionaries, values)

    Code as Data Source: Treat the Python code as a structured document containing the ground truth data. You do not need to ”run” the code, but you must ”read” and understand the variables (lists, dictionaries, values)

    work page

  60. [60]

    Semantic Mapping: Map the terms in the question to the corresponding variables and indices in the code

    work page

  61. [61]

    ## Solution Framework Phase 1: Code Structure Analysis • Scan the code to identify key variables and data lists

    Logical Derivation: Based on the extracted data, perform the necessary logical reasoning to answer the question. ## Solution Framework Phase 1: Code Structure Analysis • Scan the code to identify key variables and data lists. • Understand the relationship between different lists (e.g.,x axis usually corresponds toy axis by index). Phase 2: Information Ext...

    work page

  62. [63]

    The last line of your response must be EXACTLY: ”Therefore, the final answer is <answer>ANSWER</answer>.” ## Input Data

    work page

  63. [64]

    Chart Code: {chart code}

    work page

  64. [65]

    Your response will be used as a high-quality example to train a new AI model

    Question: {generated question} 28 ChartVerse: Scaling Chart Reasoning via Reliable Programmatic Synthesis from Scratch Table 13: Prompt for Vision-Language Model (VLM) CoT Distillation CoT Distillation Prompt ## Role You are an expert in science and visual reasoning with advanced capabilities in multimodal analysis. Your response will be used as a high-qu...

    work page

  65. [66]

    Never ignore visual elements

    Equal Weight to All Inputs: Information from images is AS IMPORTANT as text. Never ignore visual elements. 2.Systematic Analysis: Follow a rigorous, reproducible approach. 3.Precision and Accuracy: Double-check all calculations. 4.Adaptive Reasoning: Choose the most appropriate method based on context. ## Solution Framework Phase 1: Comprehensive Informat...

    work page

  66. [67]

    The final result goes ONLY inside<answer>...</answer>tags

    work page

  67. [68]

    Include units inside the tags when required (e.g.,<answer>5.2 m/s</answer>)

    work page

  68. [69]

    The last line must be EXACTLY: ”Therefore, the final answer is<answer>ANSWER</answer>.” ## Input Data Problem: {question} (Visual Inputs are provided via the vLLM interface) 29

    work page