Pith. sign in

REVIEW 3 major objections 4 minor 2 cited by

Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that a supervised cold start before reinforcement learning makes multimodal math reasoning consistently stronger than either stage alone.

desk verdict A useful empirical study with a real overclaim: the two-stage recipe beats RL-only on average but not per-benchmark at 7B. read the letter →

arxiv 2505.22334 v2 pith:H7VCLRCO submitted 2025-05-28 cs.CL cs.AIcs.CVcs.LG

Lai Wei , Yuting Li , Kaipeng Zheng , Chen Wang , Yue Wang , Linghe Kong , Lichao Sun , Weiran Huang This is my paper · ORCID
classification cs.CLcs.AIcs.CVcs.LG
keywords multimodalreasoningreinforcementlearningcoldstartsupervisedfine-tuningchain-of-thoughtGRPOahamomentlargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the best way to make small multimodal models reason better about visual math is to teach them a strong reasoning format first, then let reinforcement learning refine it. The authors test a two-stage pipeline: supervised fine-tuning on distilled chain-of-thought solutions as a cold start, followed by GRPO reinforcement learning. Across four visual-math benchmarks, they report that the combination consistently beats either stage alone and reaches the top open-source scores at 3B and 7B scales. They also present evidence that reflective 'aha moment' phrasing already exists before reinforcement learning and does not by itself signal correct answers.

What carries the argument

The load-bearing mechanism is the cold start: supervised fine-tuning on chain-of-thought data distilled by rejection sampling from a stronger teacher, which installs a structured reasoning format for the subsequent RL stage to refine. GRPO, a policy-gradient algorithm that estimates each response's advantage from rewards normalized within a sampled group rather than from a learned value model, carries the reinforcement signal. The paper's diagnostic tool is a keyword-frequency check for reflective phrases such as 're-evaluate' and 're-check', used to argue that such patterns are neither emergent nor accuracy-bearing.

What would settle it

Score a random sample of SFT-only, RL-only, and combined-pipeline outputs by human annotators or the original benchmark answer keys and compare the three systems; if the combined pipeline's edge over both single-stage systems shrinks or vanishes, the main conclusion is an evaluation artifact.

Watch

Extended reading notes

Core claim

The paper's central claim is that in multimodal mathematical reasoning, reinforcement learning does not self-start: a supervised cold start on structured chain-of-thought data is what makes later RL gains possible. Using 3B and 7B vision-language models, the authors show that the full SFT+RL pipeline beats SFT-only and RL-only on four visual-math benchmarks, with the 7B model averaging 55.66 and the 3B model 50.84. They also claim that reflective 'aha moment' phrases appear in the base model before any RL, that RL raises their frequency without raising their accuracy, and that cold-start quality predicts the final score after RL.

Load-bearing premise

The central assumption is that the automated language-model judge used to score answers is unbiased across models; if it secretly prefers the answer format the cold start teaches, the reported gains could be an artifact of style rather than reasoning.

Editorial extensions

If this is right

  • At both 3B and 7B scales, the SFT+RL pipeline scores higher on MathVision, MathVerse, MathVista, and We-Math than the same model trained with SFT alone or RL alone.
  • Cold-start data quality predicts final results: distillation from a stronger teacher yields larger gains than weaker cold-start strategies, so the SFT investment carries through the RL stage.
  • Reflective 'aha moment' language occurs in base multimodal models before RL and its frequency can rise during RL without raising accuracy, so it should not be read as evidence of emergent reasoning.
  • Even SFT on incorrect or unjudged chain-of-thought data helps over the base model, indicating the reasoning format can be learned somewhat independently of answer correctness.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: the ordering 'high-quality SFT first, RL second' could be stress-tested on text-only reasoning and on models larger than 7B, since the paper only reports 3B and 7B scales.
  • If the automated judge is format-sensitive, a human-scored sample could show that part of the SFT+RL advantage is style alignment rather than stronger reasoning; the paper does not run that check.
  • The keyword-frequency finding suggests that other projects using reflection phrasing as an RL success signal should track answer accuracy instead, a monitoring change the paper leaves implicit.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper studies multimodal reasoning in MLLMs. Its first contribution is an empirical observation that 'aha moment' reflection patterns (identified by keyword matching of 're-evaluate'/'re-check') already occur in Qwen2.5-VL before RL training and that their presence is not positively correlated with answer accuracy. Based on this, the authors propose a two-stage post-training recipe: supervised fine-tuning (SFT) on distilled chain-of-thought data as a cold start, followed by GRPO reinforcement learning. They report state-of-the-art results among open-source 3B and 7B models on MathVision, MathVerse, MathVista, and We-Math, and provide ablations on cold-start data types, data quality, and the role of aha-moment training data. The paper includes code, detailed experimental setup, and a limitations section.

Significance. If the central claims hold, the paper provides a practical and simple recipe for improving multimodal mathematical reasoning: invest in a high-quality SFT cold start before RL, rather than relying on RL alone. The work is also potentially valuable for the community's understanding of 'aha moment' phenomena, especially the negative result that such reflective patterns are not sufficient for reasoning gains. Strengths include the release of code, experiments at two model scales, and a direct test of the aha-moment hypothesis by training on aha-moment-filtered data (Figure 4). The main caveat is that the strongest claim, that SFT+RL 'consistently outperforms' both SFT-only and RL-only, is not fully supported by the paper's own ablation table, and the evaluation protocol relies on a single LLM judge without external validation.

major comments (3)
  1. [Abstract; §7; Table 3] The abstract and §7 claim that the combined SFT+RL approach 'consistently outperforms both SFT-only and RL-only methods.' Table 3 does not support this at the 7B scale: SFT+RL scores 48.58 versus 49.29 for RL-only on MathVerse, and 73.40 versus 73.50 on MathVista. Thus RL-only is better on two of the four benchmarks, and the average improvement relies entirely on the We-Math gap. No error bars, confidence intervals, or paired significance tests are reported. Either add per-benchmark significance testing with multiple seeds or temper the claim to 'on average' / 'in most settings' throughout the paper.
  2. [Appendix A.3] The evaluation uses Qwen2.5-32B-Instruct as an LLM judge with no reported agreement against human annotations or original benchmark answer extraction. The paper notes that reported scores may differ from original papers due to protocol variations, but it does not clarify whether all baseline models in Table 1 were re-evaluated under the same protocol or whether baseline numbers are quoted from their source papers. If baselines were not re-evaluated, the SOTA comparisons in Figure 1 and Table 1 are apples-to-oranges. Please specify the evaluation protocol for each baseline model and provide a human-validated subset (e.g., 100–200 answers per benchmark) with judge agreement statistics.
  3. [§3; §6.3; Appendix C] The 'aha moment' analysis relies entirely on keyword matching for 're-evaluate' and 're-check,' which the authors acknowledge in Appendix C may introduce subjectivity or incompleteness. The direct training experiment in Figure 4 mitigates this concern to some extent, but the negative result is restricted to this narrow keyword-based definition. Please discuss how the conclusions might change under alternative, more semantic definitions of reflection (e.g., LLM-annotated or parse-based detection) and, if feasible, include a small validation of the keyword detector against manual annotation.
minor comments (4)
  1. [§3] The reference for the MathVision dataset is inconsistent: §3 cites [50], but Appendix A.3 and the reference list identify MathVision as [49].
  2. [Table 1] The rows for OpenVLThinker-7B and LMM-R1-3B report identical scores (25.30, 42.79, 64.10, 63.91, 49.03). This is likely a copy-paste error; please verify and correct.
  3. [Appendix C] There is a typo: 'currectly' should be 'currently.'
  4. [Figure 4] The legend reads 'Base Model Data with Aha Moment Randomly Selected Data,' which is ambiguous. Please specify which color/line corresponds to which condition.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the two-stage recipe and the "aha moment" findings are empirical, not derived from their own assumptions or from self-citations.

full rationale

The paper's central claim (SFT cold start + GRPO outperforms SFT-only and RL-only) is supported by direct ablations in Tables 3 and 4, not by fitting a parameter and then predicting it back. The "aha moment already exists" observation is a keyword-frequency analysis of base and RL-tuned models, and the follow-up causal claim is tested by training on Reflection-CoT v2 data versus random distilled data in Figure 4, so the conclusion is not imposed by construction. The only self-reference is Ref [53] for effective rank in Table 2, which is a side analysis rather than a load-bearing premise; the metric is an external, parameter-free measure applied to activations. Appendix C's caveat that "aha moment" identification uses keyword matching is a robustness limitation, not a circularity. The main weaknesses, namely the LLM-judge evaluation in Appendix A.3 and the fact that at 7B RL-only edges SFT+RL on MathVerse and MathVista in Table 3, concern support for the "consistently outperforms" claim and evaluation validity, not circularity of the derivation.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

No numerical parameters are fit to the evaluation benchmarks; the training hyperparameters (3 epochs, 2 RL episodes, learning rates 1e-5 and 1e-6, rejection sampling budget of 24) are standard choices, not fitted to the reported results. The listed axioms are the main background assumptions the central claims rest on, especially the reliability of the LLM judge and the keyword definition of aha moments.

assumptions (4)
  • domain assumption GRPO with accuracy rewards is an effective optimization method for multimodal reasoning
    The method relies on GRPO as in DeepSeek-R1 (Section 4.2); no comparison to other RL algorithms is provided.
  • domain assumption The training data distribution (50k examples from 13 datasets) is relevant to the four evaluation benchmarks
    The RL and SFT stages use the same training dataset, and generalization to MathVision, MathVerse, MathVista, and We-Math is assumed (Section 5.1).
  • domain assumption An LLM judge (Qwen2.5-32B-Instruct) can accurately evaluate correctness of free-form math answers
    All benchmark scores are produced by this judge, and the authors note scores may differ from original papers (Appendix A.3).
  • ad hoc to paper Keyword matching of 're-evaluate' and 're-check' identifies 'aha moment' patterns
    Section 3 defines aha moments by these keywords; the authors acknowledge this heuristic in the Limitations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start." pith.science (2026). https://pith.science/paper/H7VCLRCO

@misc{pith2026250522334,
  author       = {Pith},
  title        = {Pith review of: Advancing Multimodal Reasoning via Reinforcement Learning with Cold Start},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/H7VCLRCO}},
  note         = {Machine review of arXiv:2505.22334}
}
abstract

Recent advancements in large language models (LLMs) have demonstrated impressive chain-of-thought reasoning capabilities, with reinforcement learning (RL) playing a crucial role in this progress. While "aha moment" patterns--where models exhibit self-correction through reflection--are often attributed to emergent properties from RL, we first demonstrate that these patterns exist in multimodal LLMs (MLLMs) prior to RL training but may not necessarily correlate with improved reasoning performance. Building on these insights, we present a comprehensive study on enhancing multimodal reasoning through a two-stage approach: (1) supervised fine-tuning (SFT) as a cold start with structured chain-of-thought reasoning patterns, followed by (2) reinforcement learning via GRPO to further refine these capabilities. Our extensive experiments show that this combined approach consistently outperforms both SFT-only and RL-only methods across challenging multimodal reasoning benchmarks. The resulting models achieve state-of-the-art performance among open-source MLLMs at both 3B and 7B scales, with our 7B model showing substantial improvements over base models (e.g., 66.3 %$\rightarrow$73.4 % on MathVista, 62.9 %$\rightarrow$70.4 % on We-Math) and our 3B model achieving performance competitive with several 7B models. Overall, this work provides practical guidance for building advanced multimodal reasoning models. Our code is available at https://github.com/waltonfuture/RL-with-Cold-Start.

Figures

Figures reproduced from arXiv: 2505.22334 by the authors.

Figure 1
Figure 1. Performance comparison between our models and other advanced models on different [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The frequency and accuracy of models’ responses with and without “aha moment”. The [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Method overview. Our approach consists of two stages: (1) a cold start phase using [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison of model performance when trained on data with “aha moment” patterns [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

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

  1. LaRe: Latent Refocusing for Multimodal Reasoning

    cs.CV 2025-11 reject novelty 6.0 of 10

    LaRe performs iterative visual refocusing in latent space and reports accuracy gains with fewer tokens, but its main experiments compare against baselines trained with less data.

  2. SLMs as Multi-Agent Routers: A Progressive SFT and Reinforcement Learning Approach

    cs.CL 2026-07 reject novelty 5.0 of 10

    A 0.6B router trained by SFT+RL on retrieval-quality rewards reaches 0.771 NDCG@10 across 11 agents, beating intent-prompted LLMs and cutting latency by 82%.

Reference graph

Works this paper leans on

91 extracted references · 28 canonical work pages · cited by 2 Pith papers

  1. [1]

    The claude 3 model family: Opus, sonnet, haiku

    Anthropic. The claude 3 model family: Opus, sonnet, haiku. https://www-cdn.anthropic.com/ de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf, 2024. Preprint

    2024

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  3. [3]

    An augmented benchmark dataset for geometric question answering through dual parallel text encoding

    Jie Cao and Jing Xiao. An augmented benchmark dataset for geometric question answering through dual parallel text encoding. In Proceedings of the 29th international conference on computational linguistics , pages 1511–1520, 2022

    2022

  4. [4]

    Sft or rl? an early investigation into training r1-like reasoning large vision-language models

    Hardy Chen, Haoqin Tu, Fali Wang, Hui Liu, Xianfeng Tang, Xinya Du, Yuyin Zhou, and Cihang Xie. Sft or rl? an early investigation into training r1-like reasoning large vision-language models. https://github.com/UCSC-VLAA/VLAA-Thinking , 2025

    2025

  5. [5]

    Geoqa: A geometric question answering benchmark towards multimodal numerical reasoning

    Jiaqi Chen, Jianheng Tang, Jinghui Qin, Xiaodan Liang, Lingbo Liu, Eric Xing, and Liang Lin. Geoqa: A geometric question answering benchmark towards multimodal numerical reasoning. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021 , pages 513–523, 2021

    2021

  6. [6]

    R1-v: Reinforcing super generalization ability in vision-language models with less than $3

    Liang Chen, Lei Li, Haozhe Zhao, Yifan Song, and Vinci. R1-v: Reinforcing super generalization ability in vision-language models with less than $3. https://github.com/Deep-Agent/R1-V , 2025. Accessed: 2025-02-02

    2025

  7. [7]

    Sharegpt4v: Improving large multi-modal models with better captions

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. In European Conference on Computer Vision, pages 370–387. Springer, 2024. 10

    2024

  8. [8]

    M3cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought

    Qiguang Chen, Libo Qin, Jin Zhang, Zhi Chen, Xiao Xu, and Wanxiang Che. M3cot: A novel benchmark for multi-domain multi-step multi-modal chain-of-thought. arXiv preprint arXiv:2405.16473, 2024

    arXiv 2024

Show all 91 references
  1. [9]

    Vision-language models can self-improve reasoning via reflection

    Kanzhi Cheng, Yantao Li, Fangzhi Xu, Jianbing Zhang, Hao Zhou, and Yang Liu. Vision-language models can self-improve reasoning via reflection. arXiv preprint arXiv:2411.00855, 2024

    2024 arXiv

  2. [10]

    Openvlthinker: An early exploration to complex vision-language reasoning via iterative self-improvement

    Yihe Deng, Hritik Bansal, Fan Yin, Nanyun Peng, Wei Wang, and Kai-Wei Chang. Openvlthinker: An early exploration to complex vision-language reasoning via iterative self-improvement. arXiv preprint arXiv:2503.17352, 2025

    2025 arXiv

  3. [11]

    Insight-v: Exploring long-chain visual reasoning with multimodal large language models

    Yuhao Dong, Zuyan Liu, Hai-Long Sun, Jingkang Yang, Winston Hu, Yongming Rao, and Ziwei Liu. Insight-v: Exploring long-chain visual reasoning with multimodal large language models. arXiv preprint arXiv:2411.14432, 2024

    2024 arXiv

  4. [12]

    Text-to-sql empowered by large language models: A benchmark evaluation

    Dawei Gao, Haibin Wang, Yaliang Li, Xiuyu Sun, Yichen Qian, Bolin Ding, and Jingren Zhou. Text-to-sql empowered by large language models: A benchmark evaluation. arXiv preprint arXiv:2308.15363, 2023

    2023 arXiv

  5. [13]

    Introducing gemini 2.0: our new ai model for the agentic era, 2024

    Google. Introducing gemini 2.0: our new ai model for the agentic era, 2024. URL https: //blog.google/technology/google-deepmind/google-gemini-ai-update-december-2024/ #ceo-message

    2024

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  7. [15]

    Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale

    Jarvis Guo, Tuney Zheng, Yuelin Bai, Bo Li, Yubo Wang, King Zhu, Yizhi Li, Graham Neubig, Wenhu Chen, and Xiang Yue. Mammoth-vl: Eliciting multimodal reasoning with instruction tuning at scale. arXiv preprint arXiv:2412.05237, 2024

    2024 arXiv

  8. [16]

    Infimm-webmath-40b: Advancing multimodal pre-training for enhanced mathematical reasoning

    Xiaotian Han, Yiren Jian, Xuefeng Hu, Haogeng Liu, Yiqi Wang, Qihang Fan, Yuang Ai, Huaibo Huang, Ran He, Zhenheng Yang, et al. Infimm-webmath-40b: Advancing multimodal pre-training for enhanced mathematical reasoning. In The 4th Workshop on Mathematical Reasoning and AI at Ne...

    2024

  9. [17]

    Visual sketchpad: Sketching as a visual chain of thought for multimodal language models

    Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024

    2024 arXiv

  10. [18]

    Vision-r1: Incentivizing reasoning capability in multimodal large language models

    Wenxuan Huang, Bohan Jia, Zijie Zhai, Shaosheng Cao, Zheyu Ye, Fei Zhao, Zhe Xu, Yao Hu, and Shaohui Lin. Vision-r1: Incentivizing reasoning capability in multimodal large language models. arXiv preprint arXiv:2503.06749, 2025

    2025 arXiv

  11. [19]

    Gpt-4o system card

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024

    2024 arXiv

  12. [20]

    Openai o1 system card

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Hel- yar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card. arXiv preprint arXiv:2412.16720, 2024

    2024 arXiv

  13. [21]

    Figureqa: An annotated figure dataset for visual reasoning

    Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017

    2017 arXiv

  14. [22]

    A diagram is worth a dozen images

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. InComputer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14 , pages 235–251. ...

    2016

  15. [23]

    Textbook question answering with multi-modal context graph understanding and self-supervised open-set comprehension

    Daesik Kim, Seonhoon Kim, and Nojun Kwak. Textbook question answering with multi-modal context graph understanding and self-supervised open-set comprehension. arXiv preprint arXiv:1811.00232, 2018

    2018 arXiv

  16. [24]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024

    2024 arXiv

  17. [25]

    Llms can easily learn to reason from demonstrations structure, not content, is what matters! arXiv preprint arXiv:2502.07374, 2025

    Dacheng Li, Shiyi Cao, Tyler Griggs, Shu Liu, Xiangxi Mo, Eric Tang, Sumanth Hegde, Kourosh Hakhamaneshi, Shishir G Patil, Matei Zaharia, et al. Llms can easily learn to reason from demonstrations structure, not content, is what matters! arXiv preprint arXiv:2502.07374, 2025. 11

    2025 arXiv

  18. [26]

    Small models struggle to learn from strong reasoners

    Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubra- manian, and Radha Poovendran. Small models struggle to learn from strong reasoners. arXiv preprint arXiv:2502.12143, 2025

    2025

  19. [27]

    Clevr-math: A dataset for compositional language, visual and mathematical reasoning

    Adam Dahlgren Lindström and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358, 2022

    2022 arXiv

  20. [28]

    Inter- gps: Interpretable geometry problem solving with formal language and symbolic reasoning

    Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. Inter- gps: Interpretable geometry problem solving with formal language and symbolic reasoning. arXiv preprint arXiv:2105.04165, 2021

    2021 arXiv

  21. [29]

    Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning

    Pan Lu, Liang Qiu, Jiaqi Chen, Tony Xia, Yizhou Zhao, Wei Zhang, Zhou Yu, Xiaodan Liang, and Song-Chun Zhu. Iconqa: A new benchmark for abstract diagram understanding and visual language reasoning. arXiv preprint arXiv:2110.13214, 2021

    arXiv

  22. [30]

    Learn to explain: Multimodal reasoning via thought chains for science question answering

    Pan Lu, Swaroop Mishra, Tanglin Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering. Advances in Neural Information Processing Systems , 35:2507–2...

    2022

  23. [31]

    Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning

    Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. arXiv preprint arXiv:2209.14610, 2022

    2022 arXiv

  24. [32]

    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

    arXiv

  25. [33]

    Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct

    Haipeng Luo, Qingfeng Sun, Can Xu, Pu Zhao, Jianguang Lou, Chongyang Tao, Xiubo Geng, Qingwei Lin, Shifeng Chen, and Dongmei Zhang. Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct. arXiv preprint arXiv:2308.09583, 2023

    2023 arXiv

  26. [34]

    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. arXiv preprint arXiv:2203.10244, 2022

    2022 arXiv

  27. [35]

    Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning, 2025

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Botian Shi, Wenhai Wang, Junjun He, Kaipeng Zhang, Ping Luo, Yu Qiao, Qiaosheng Zhang, and Wenqi Shao. Mm-eureka: Exploring visual aha moment with rule-based large-scale reinforcement learning, 20...

    2025

  28. [36]

    Multimath: Bridging visual and mathematical reasoning for large language models

    Shuai Peng, Di Fu, Liangcai Gao, Xiuqin Zhong, Hongguang Fu, and Zhi Tang. Multimath: Bridging visual and mathematical reasoning for large language models. arXiv preprint arXiv:2409.00147, 2024

    2024 arXiv

  29. [37]

    Skywork r1v: Pioneering multimodal reasoning with chain-of-thought

    Yi Peng, Xiaokun Wang, Yichen Wei, Jiangbo Pei, Weijie Qiu, Ai Jian, Yunzhuo Hao, Jiachun Pan, Tianyidan Xie, Li Ge, et al. Skywork r1v: Pioneering multimodal reasoning with chain-of-thought. arXiv preprint arXiv:2504.05599, 2025

    2025 arXiv

  30. [38]

    We-math: Does your large multimodal model achieve human-like mathematical reasoning? arXiv preprint arXiv:2407.01284, 2024

    Runqi Qiao, Qiuna Tan, Guanting Dong, Minhui Wu, Chong Sun, Xiaoshuai Song, Zhuoma GongQue, Shanglin Lei, Zhe Wei, Miaoxuan Zhang, et al. We-math: Does your large multimodal model achieve human-like mathematical reasoning? arXiv preprint arXiv:2407.01284, 2024

    2024 arXiv

  31. [39]

    Solving geometry problems: Combining text and diagram interpretation

    Minjoon Seo, Hannaneh Hajishirzi, Ali Farhadi, Oren Etzioni, and Clint Malcolm. Solving geometry problems: Combining text and diagram interpretation. In Proceedings of the 2015 conference on empirical methods in natural language processing, pages 1466–1476, 2015

    2015

  32. [40]

    Rethinking reflection in pre-training

    Darsh J Shah, Peter Rushton, Somanshu Singla, Mohit Parmar, Kurt Smith, Yash Vanjani, Ashish Vaswani, Adarsh Chaluvaraju, Andrew Hojel, Andrew Ma, et al. Rethinking reflection in pre-training. arXiv preprint arXiv:2504.04022, 2025

    2025 arXiv

  33. [41]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  34. [42]

    Vlm-r1: A stable and generalizable r1-style large vision-language model

    Haozhan Shen, Zilun Zhang, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. Vlm-r1: A stable and generalizable r1-style large vision-language model. https://github.com/om-ai-lab/VLM-R1 , 2025. Accessed: 2025-02-15

    2025

  35. [43]

    Math-llava: Bootstrapping mathematical reasoning for multimodal large language models

    Wenhao Shi, Zhiqiang Hu, Yi Bin, Junhua Liu, Yang Yang, See-Kiong Ng, Lidong Bing, and Roy Ka-Wei Lee. Math-llava: Bootstrapping mathematical reasoning for multimodal large language models. arXiv preprint arXiv:2406.17294, 2024. 12

    2024 arXiv

  36. [44]

    Qvq: To see the world with wisdom, December 2024

    Qwen Team. Qvq: To see the world with wisdom, December 2024. URL https://qwenlm.github.io/ blog/qvq-72b-preview/

    2024

  37. [45]

    Qwq: Reflect deeply on the boundaries of the unknown, November 2024

    Qwen Team. Qwq: Reflect deeply on the boundaries of the unknown, November 2024. URL https: //qwenlm.github.io/blog/qwq-32b-preview/

    2024

  38. [46]

    Llamav-o1: Rethinking step-by-step visual reasoning in llms

    Omkar Thawakar, Dinura Dissanayake, Ketan More, Ritesh Thawkar, Ahmed Heakl, Noor Ahsan, Yuhao Li, Mohammed Zumri, Jean Lahoud, Rao Muhammad Anwer, et al. Llamav-o1: Rethinking step-by-step visual reasoning in llms. arXiv preprint arXiv:2501.06186, 2025

    2025 arXiv

  39. [47]

    Dart-math: Difficulty-aware rejection tuning for mathematical problem-solving

    Yuxuan Tong, Xiwen Zhang, Rui Wang, Ruidong Wu, and Junxian He. Dart-math: Difficulty-aware rejection tuning for mathematical problem-solving. Advances in Neural Information Processing Systems , 37:7821–7846, 2024

    2024

  40. [48]

    Llama 2: Open foundation and fine-tuned chat models

    Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023

    2023 arXiv

  41. [49]

    Measuring multimodal mathematical reasoning with math-vision dataset

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hong- sheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37:95095–95169, 2025

    2025

  42. [50]

    Math-shepherd: Verify and reinforce llms step-by-step without human annotations

    Peiyi Wang, Lei Li, Zhihong Shao, Runxin Xu, Damai Dai, Yifei Li, Deli Chen, Yu Wu, and Zhifang Sui. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (V o...

    2024

  43. [51]

    Sota with less: Mcts-guided sample selection for data-efficient visual reasoning self-improvement

    Xiyao Wang, Zhengyuan Yang, Chao Feng, Hongjin Lu, Linjie Li, Chung-Ching Lin, Kevin Lin, Furong Huang, and Lijuan Wang. Sota with less: Mcts-guided sample selection for data-efficient visual reasoning self-improvement. arXiv preprint arXiv:2504.07934, 2025

    2025 arXiv

  44. [52]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824–24837, 2022

    2022

  45. [53]

    Large language model evaluation via matrix entropy

    Lai Wei, Zhiquan Tan, Chenghai Li, Jindong Wang, and Weiran Huang. Large language model evaluation via matrix entropy. arXiv preprint arXiv:2401.17139, 2024

    2024 arXiv

  46. [54]

    Thinkpatterns-21k: A systematic study on the impact of thinking patterns in llms

    Pengcheng Wen, Jiaming Ji, Chi-Min Chan, Juntao Dai, Donghai Hong, Yaodong Yang, Sirui Han, and Yike Guo. Thinkpatterns-21k: A systematic study on the impact of thinking patterns in llms. arXiv preprint arXiv:2503.12918, 2025

    2025

  47. [55]

    Llava-cot: Let vision language models reason step-by-step

    Guowei Xu, Peng Jin, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. Llava-cot: Let vision language models reason step-by-step. CoRR, abs/2411.10440, 2024

    2024 arXiv

  48. [56]

    Tbac-vlr1-3b-preview, 2025

    Junzhe Xu and Yuyang yin. Tbac-vlr1-3b-preview, 2025. URL https://huggingface.co/ TencentBAC/TBAC-VLR1-3B-preview

    2025

  49. [57]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115, 2024

    2024 arXiv

  50. [58]

    Mathglm-vision: Solving mathematical problems with multi-modal large language model

    Zhen Yang, Jinhao Chen, Zhengxiao Du, Wenmeng Yu, Weihan Wang, Wenyi Hong, Zhihuan Jiang, Bin Xu, Yuxiao Dong, and Jie Tang. Mathglm-vision: Solving mathematical problems with multi-modal large language model. arXiv preprint arXiv:2409.13729, 2024

    2024 arXiv

  51. [59]

    Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search

    Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, et al. Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search. arXiv preprint arXiv:2412.18319, 2024

    2024 arXiv

  52. [60]

    Easyr1: An efficient, scalable, multi-modality rl training framework

    Zheng Yaowei, Lu Junting, Wang Shenzhi, Feng Zhangchi, Kuang Dongdong, and Xiong Yuwen. Easyr1: An efficient, scalable, multi-modality rl training framework. https://github.com/hiyouga/EasyR1, 2025

    2025

  53. [61]

    Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl, 2025

    Peng Yingzhe, Zhang Gongrui, Zhang Miaosen, You Zhiyuan, Liu Jie, Zhu Qipeng, Yang Kai, Xu Xingzhong, Geng Xin, and Yang Xu. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl, 2025

    2025

  54. [62]

    Perception-r1: Pioneering perception policy with reinforcement learning

    En Yu, Kangheng Lin, Liang Zhao, Jisheng Yin, Yana Wei, Yuang Peng, Haoran Wei, Jianjian Sun, Chunrui Han, Zheng Ge, et al. Perception-r1: Pioneering perception policy with reinforcement learning. arXiv preprint arXiv:2504.07954, 2025. 13

    2025 arXiv

  55. [63]

    Metamath: Bootstrap your own mathematical questions for large language models

    Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284, 2023

    2023 arXiv

  56. [64]

    Llama-berry: Pairwise optimization for o1-like olympiad-level mathematical reasoning

    Di Zhang, Jianbo Wu, Jingdi Lei, Tong Che, Jiatong Li, Tong Xie, Xiaoshui Huang, Shufei Zhang, Marco Pavone, Yuqiang Li, et al. Llama-berry: Pairwise optimization for o1-like olympiad-level mathematical reasoning. arXiv preprint arXiv:2410.02884, 2024

    arXiv

  57. [65]

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

    2024

  58. [66]

    Mavis: Mathematical visual instruction tuning with an automatic data engine

    Renrui Zhang, Xinyu Wei, Dongzhi Jiang, Ziyu Guo, Shicheng Li, Yichi Zhang, Chengzhuo Tong, Jiaming Liu, Aojun Zhou, Bin Wei, et al. Mavis: Mathematical visual instruction tuning with an automatic data engine. arXiv preprint arXiv:2407.08739, 2024

    2024 arXiv

  59. [67]

    Improve vision language model chain-of-thought reasoning

    Ruohong Zhang, Bowen Zhang, Yanghao Li, Haotian Zhang, Zhiqing Sun, Zhe Gan, Yinfei Yang, Ruoming Pang, and Yiming Yang. Improve vision language model chain-of-thought reasoning. arXiv preprint arXiv:2410.16198, 2024

    2024 arXiv

  60. [68]

    Question-guided knowledge graph re-scoring and injection for knowledge graph question answering

    Yu Zhang, Kehai Chen, Xuefeng Bai, Zhao Kang, Quanjiang Guo, and Min Zhang. Question-guided knowledge graph re-scoring and injection for knowledge graph question answering. In Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen, editors, Findings of the Association for Computati...

    2024 doi

  61. [69]

    Swift:a scalable lightweight infrastructure for fine-tuning, 2024

    Yuze Zhao, Jintao Huang, Jinghan Hu, Xingjun Wang, Yunlin Mao, Daoze Zhang, Zeyinzi Jiang, Zhikai Wu, Baole Ai, Ang Wang, Wenmeng Zhou, and Yingda Chen. Swift:a scalable lightweight infrastructure for fine-tuning, 2024. URL https://arxiv.org/abs/2408.05517

    2024 arXiv

  62. [70]

    aha moment

    Hengguang Zhou, Xirui Li, Ruochen Wang, Minhao Cheng, Tianyi Zhou, and Cho-Jui Hsieh. R1-zero’s "aha moment" in visual reasoning on a 2b non-sft model, 2025. URL https://arxiv.org/abs/2503. 05132

    2025

  63. [71]

    Math-puma: Progressive upward multimodal alignment to enhance mathematical reasoning

    Wenwen Zhuang, Xin Huang, Xiantao Zhang, and Jin Zeng. Math-puma: Progressive upward multimodal alignment to enhance mathematical reasoning. arXiv preprint arXiv:2408.08640, 2024. 14 Appendix A More Implementation Details A.1 Prompts We provide the prompt used to synthesize Se...

    2024 arXiv

  64. [72]

    Draft Response: Generate an initial response

  65. [73]

    aha moment

    Critical Comments: Analyze your draft response by considering: • Potential weaknesses or gaps • Logical flaws or inconsistencies • Missing perspectives or alternatives • Areas for improvement • Suggestions for a better version • Steering toward the given answer The critical co...

  66. [74]

    Opposite sides are parallel

  67. [75]

    Opposite angles are equal

  68. [76]

    Given: The sum of angle B and angle D is 100◦

    Consecutive angles are supplementary (they add up to 180◦). Given: The sum of angle B and angle D is 100◦. Since AD ∥ BC and AB ∥ CD, angles B and D are consecutive angles. Therefore, they are supplementary, meaning they add up to 180◦. So, we can write: angle B + angle D = 18...

  69. [77]

    • Diameter BE of circle O means that BE is a straight line passing through the center of the circle

    Identify the given information: • Quadrilateral ABCD is inscribed in circle O. • Diameter BE of circle O means that BE is a straight line passing through the center of the circle. • AE is connected to B and E. • ∠BCD = 2× ∠BAD

  70. [78]

    Therefore, ∠BAD + ∠BCD = 180◦

    Use the property of cyclic quadrilaterals: In a cyclic quadrilateral, the sum of the opposite angles is 180◦. Therefore, ∠BAD + ∠BCD = 180◦

  71. [79]

    Express ∠BCD in terms of ∠BAD: Given ∠BCD = 2× ∠BAD, substitute this into the equation: ∠BAD + 2× ∠BAD = 180◦, 3 × ∠BAD = 180◦, ∠BAD = 60◦, ∠BCD = 2× 60◦ = 120◦

  72. [80]

    Determine ∠ADE: Since BE is a diameter, ∠ADE is a right angle because the angle subtended by a diameter in a semicircle is 90◦: ∠ADE = 90◦. 17

  73. [81]

    The sum of the angles in triangle ADE is 180◦: ∠DAE + ∠ADE + ∠AED = 180◦, ∠DAE + 90◦ + ∠AED = 180◦, ∠DAE + ∠AED = 90◦

    Determine ∠DAE: ∠DAE is part of triangle ADE, where ∠ADE = 90◦. The sum of the angles in triangle ADE is 180◦: ∠DAE + ∠ADE + ∠AED = 180◦, ∠DAE + 90◦ + ∠AED = 180◦, ∠DAE + ∠AED = 90◦

  74. [82]

    Determine ∠AED: ∠AED is the same as ∠BCD because they subtend the same arc CD: ∠AED = 120◦

  75. [83]

    Since ∠DAE cannot be negative, we must re-evaluate the problem

    Solve for ∠DAE: Substitute ∠AED = 120◦ into the equation: ∠DAE + 120◦ = 90◦, ∠DAE = −30◦. Since ∠DAE cannot be negative, we must re-evaluate the problem. The correct interpretation is that ∠AED = 60◦ and not 120◦, because ∠AED does not subtend the same arc as ∠BCD . Instead, t...

  76. [84]

    • Given the perimeter is 30, we can find the length of one side by dividing the perimeter by 3: Side length = 30 3 = 10

    Calculate the side length of the equilateral triangle: • The perimeter of an equilateral triangle is the sum of the lengths of its three sides. • Given the perimeter is 30, we can find the length of one side by dividing the perimeter by 3: Side length = 30 3 = 10

  77. [85]

    • In a 30-60-90 triangle, the ratio of the sides opposite the 30◦, 60◦, and 90◦ angles is 1 : √ 3 : 2

    Use the properties of an equilateral triangle: • In an equilateral triangle, the altitude also bisects the base, creating two 30-60-90 right triangles. • In a 30-60-90 triangle, the ratio of the sides opposite the 30◦, 60◦, and 90◦ angles is 1 : √ 3 : 2

  78. [86]

    • The side opposite the 30◦ angle (which is half the base) is 5 (since the base is 10 and it is bisected)

    Determine the length of the altitude h: • The altitude h is the side opposite the 60◦ angle in the right triangle. • The side opposite the 30◦ angle (which is half the base) is 5 (since the base is 10 and it is bisected). • Using the ratio 1 : √ 3 : 2, the altitude h is: h = 5...

  79. [87]

    • We need to find the length of the altitude h of this triangle

    Understand the Problem: • We are given an equilateral triangle △ABC with a perimeter of 30. • We need to find the length of the altitude h of this triangle. 18

  80. [88]

    • Let the side length of the triangle be s

    Find the Side Length of the Triangle: • An equilateral triangle has all three sides of equal length. • Let the side length of the triangle be s. • The perimeter of the triangle is given by: Perimeter = s + s + s = 3s • We are given that the perimeter is 30: 3s = 30 • Solving f...

  81. [89]

    Use the Formula for the Altitude of an Equilateral Triangle: • The altitude h of an equilateral triangle with side length s is given by the formula: h = √ 3 2 s • Substitute s = 10into the formula: h = √ 3 2 · 10 • Simplify the expression: h = 5 √ 3

  82. [90]

    • In an equilateral triangle, the altitude bisects the base, creating two 30-60-90 right triangles

    Verify the Result: • The formula for the altitude of an equilateral triangle is derived from the Pythagorean theorem. • In an equilateral triangle, the altitude bisects the base, creating two 30-60-90 right triangles. • In a 30-60-90 triangle, the sides are in the ratio 1 : √ ...

  83. [91]

    5 √ 3 19

    Select the Correct Answer: • The length of the altitude h is 5 √ 3. 5 √ 3 19

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.