Pith. sign in

REVIEW 4 major objections 6 minor 75 references

Multimodal Mathematical Reasoning with Diverse Solving Perspective

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Training on multiple solution paths lifts multimodal math accuracy to 70.4%

desk verdict The MathVista gain is credible and the dataset is useful, but the paper never isolates diverse perspectives from simply more CoT data, so its central causal claim is unproven. read the letter →

arxiv 2507.02804 v2 pith:MFQJGQMC submitted 2025-07-03 cs.CL

Wenhao Shi , Zhiqiang Hu , Yi Bin , Guoqing Wang , Xing Xu , Yang Yang , See-Kiong Ng This is my paper
classification cs.CL
keywords multimodalmathematicalreasoningdiverseperspectiveschain-of-thoughtsyntheticsolutiongenerationGRPOrewarddesignMathV-DPdiversity
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 single-solution supervision is a bottleneck for multimodal math reasoning; models improve when they see several valid ways to solve the same image-question pair, plus reflective wrong attempts. To test this, the authors build MathV-DP, a dataset in which each question comes with two correct and two incorrect reasoning trajectories synthesized by a strong reasoning LLM. They then post-train Qwen2.5-VL through supervised fine-tuning followed by GRPO with rewards for correct final answers, required think-tags, and the ability to tell solutions apart and prefer correct ones. The resulting model, Qwen2.5-VL-DP, reaches 70.4% on MathVista minitest versus 68.2% for the base, 26.9% versus 25.0% on Math-V, and higher pairwise semantic diversity at K=3, 5, and 10. If correct, this is evidence that diversity-aware reward design can be as important as answer accuracy in multimodal reasoning.

What carries the argument

The central object is MathV-DP, a dataset built from MultiMath-300K samples in which DeepSeek-R1 generates two correct and two incorrect reflective solutions per image-question pair. The pairs are organized into three formats: correct solutions for SFT, pairs of different correct solutions for the discrimination reward, and pairs of one correct and one incorrect solution for the preference reward. The training machinery is GRPO with three rule-based reward functions—accuracy, think-format, and discrimination/preference—applied after supervised fine-tuning on the correct solutions.

What would settle it

Take a random sample of MathV-DP and have independent human annotators or a separate strong verifier relabel each solution as correct or incorrect and mark near-duplicate pairs. If label disagreement is substantial or many "distinct" solutions are paraphrases of one approach, retrain with corrected labels; the claimed accuracy and diversity gains should shrink or disappear if the diversity signal is the real driver.

Watch

Extended reading notes

Core claim

The central discovery claimed is that post-training with multiple distinct solution trajectories per visual math question—including deliberately incorrect ones—improves both how often the model answers correctly and how varied its reasoning is. The paper reports that Qwen2.5-VL-DP outperforms its Qwen2.5-VL-7B base by 2.2 points on MathVista minitest (70.4 vs 68.2) and by 1.9 points on Math-V (26.9 vs 25.0), with a 10-run average gain of 2.5% (±0.9%) and a 95% confidence interval of (1.94%, 3.06%). The method also raises effective semantic diversity from 36.89 to 41.58 at K=10 on the same base. Ablations show both the discrimination reward (recognizing two correct solutions as different) and the preference reward (choosing correct over incorrect) contribute, and that SFT before RL matters. The paper's stated conclusion is that moving from one-to-one image-text supervision to multiple solving perspectives is what drives the gain.

Load-bearing premise

The entire training signal depends on labels that come from the synthesizing model: if a substantial share of the automatically generated "correct" and "incorrect" labels are wrong, or if the supposedly distinct solutions are really paraphrases of one approach, both the SFT targets and the reward signals would be teaching the model something false.

Editorial extensions

If this is right

  • A post-training pipeline built on synthetic diverse solutions can lift open-source 7B multimodal models to accuracy comparable with much larger closed models on MathVista minitest.
  • Adding discrimination and preference rewards on top of accuracy and format rewards improves both accuracy and solution diversity; removing either reward degrades both benchmarks.
  • SFT followed by GRPO outperforms either stage alone, while applying RL without prior SFT is suboptimal.
  • The gains transfer beyond math: the same post-trained model improves MMMU accuracy from 58.6 to 59.4 on the Qwen2.5-VL base and from 47.8 to 49.4 on Qwen2-VL.
  • Generative diversity, measured as pairwise semantic distance over K sampled responses, consistently increases for the full model at K=3, 5, and 10.

Reading between the lines

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

  • A natural next experiment, not run in the paper, is to vary the number of correct solutions per question and measure the saturation point of both accuracy and diversity gains; the paper fixes two correct and two incorrect solutions.
  • The discrimination reward may implicitly teach the model to attend to solution structure rather than content; a permuted-pair or cross-order evaluation would test whether the reward is truly reasoning-based.
  • The same data-synthesis and reward recipe could be ported to other multimodal reasoning domains with checkable answers, such as chart QA or geometry proof, wherever multiple valid solution paths exist.
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

4 major / 6 minor

Summary. The paper introduces MathV-DP, a dataset of multiple correct and incorrect chain-of-thought solution trajectories per image-question pair, synthesized by prompting DeepSeek-R1 with formal image captions from MultiMath-300K. The authors propose a two-stage post-training recipe for 7B Qwen-VL models: supervised fine-tuning on the correct reflection-wrapped solutions, followed by GRPO with accuracy, think-format, and discrimination/preference rewards. On MathVista minitest, Qwen2.5-VL-DP reaches 70.4% vs. 68.2% for the base model, with a reported 95% CI for the average gain of (1.94%, 3.06%) over 10 inference runs; on Math-V it reaches 26.9% vs. 25.0%. The paper also reports improved generative diversity (Table III), reward ablations (Table IV), and generalization to MMMU (Table V). The main claim is that learning from diverse solving perspectives significantly improves both accuracy and diversity.

Significance. If the causal claim is established, the contribution is useful: a synthetic dataset with multiple correct/incorrect trajectories plus a reward design that rewards both correctness and diversity could be a practical recipe for post-training MLLMs for mathematical reasoning. The paper ships several strong methodological elements: a repeated-run experiment with a 95% CI on the main MathVista gain, a concrete data-generation pipeline, and open benchmarks for evaluation. However, the central attribution to diversity is not yet supported by the experiments, and the label-quality and significance-testing gaps are real. The dataset and framework are potentially reusable, so a revised version with the missing controls would be a meaningful contribution.

major comments (4)
  1. [V-D, Figure 4; V-C, Table IV] The central causal claim that diverse solving perspectives drive the gains is not supported by the current experimental design. The paper never compares MathV-DP against a control that holds the number of training solution trajectories fixed while varying their diversity (e.g., the same volume of single-perspective or paraphrase-level CoT data). The ablation in Figure 4 shows that SFT alone already improves over the base model (+1.4 points for Qwen2-VL and +1.4 for Qwen2.5-VL), while Table IV shows that removing both RL rewards still leaves +1.7 and +0.2 on MathVista and Math-V over the base. These gains could be explained by additional reasoning traces, format regularization, or data quantity rather than by diversity of perspectives. Please add a single-perspective control (same dataset size, same model, same SFT/RL protocol) and report the comparison on both benchmarks.
  2. [III-A; V-E] The correctness labels of the synthesized solutions are produced by DeepSeek-R1 and are used both as SFT targets and as the ground truth for the discrimination/preference rewards. The human validation in Section V-E covers only 1,000 samples, reports a single aggregate score per dimension without inter-annotator agreement, and does not clearly map the three reported values (0.82, 0.97, 0.95) to the three stated evaluation dimensions (overall correctness, distinction between correct and incorrect, and output distinction). The GPT-4o filter is described without its criteria. If a non-trivial fraction of the labels is wrong, the RL reward loop can actively reinforce incorrect solutions. Please report per-dimension human scores with IAA (e.g., Fleiss' kappa), a per-type label accuracy (correct vs. incorrect), and a sensitivity analysis comparing models trained on the 38K filtered subset vs. the full 40K set.
  3. [IV-B; V-B, Table III] The diversity metric depends on an unspecified dissimilarity threshold in dsem, and the reported diversity scores have no variance or significance tests. The gaps in Table III (e.g., Qwen2.5-VL-SFT at 39.46 vs. Qwen2.5-VL-DP at 40.42 for Diver@3) are small relative to typical run-to-run variation on a 1,000-sample benchmark. Please state the threshold used in dsem, report confidence intervals or paired tests over repeated generation seeds, and justify the claim that the full model is significantly more diverse than the SFT model.
  4. [V-A, Table II; V-C, Table IV] The Math-V gain (25.0 to 26.9) and the reward-ablation differences in Table IV (0.2-1.7 points) are not accompanied by any significance testing. Given that the MathVista main gain has a 95% CI of (1.94%, 3.06%), the Math-V gain of 1.9 points and the ablation differences are plausibly within noise. Please provide paired significance tests (e.g., McNemar or bootstrap over items) for the Math-V comparison and for the reward ablations, and report per-seed variation for Table IV.
minor comments (6)
  1. [Throughout] The model name is inconsistent: the Abstract and Section VI use 'Qwen-VL-DP' while the tables and Section V use 'Qwen2-VL-DP' and 'Qwen2.5-VL-DP'. Please standardize the notation to distinguish the model family and the post-trained variants.
  2. [Eqs. (5)-(6)] The sign convention in Eq. (6) is confusing: the KL term appears inside the expectation with a minus sign, whereas standard GRPO formulations place the KL penalty outside the expectation as a positive regularizer in the loss to be minimized. Clarify the expression and state explicitly whether LRL is minimized or maximized.
  3. [Figure 1] The inline text in Figure 1 is run together and nearly unreadable in the extracted PDF; the figure should be typeset with proper spacing so that the question, correct solutions, and incorrect solution are visually separate.
  4. [V-E] The notation '0.82 (±0.8%), 0.97 (±0.4%), and 0.95 (±0.5%)' is unclear: it is not stated whether the parenthetical is a standard deviation, a standard error, or a percentage-point margin, and the three dimensions are not identified. Please report each dimension with its confidence interval or dispersion measure.
  5. [V-A] The relationship between the reported 2.2% improvement (single run, Table I) and the average improvement of 2.5% (±0.9%) over 10 runs should be clarified; report the variance across runs and whether the tables use one representative run or an average.
  6. [IV-B, Eq. (7)] Equation (7) renders as '1 K 2' in the submitted PDF; this should be the binomial coefficient binom(K,2). The diversity metric citation [72] should also be verified against the precise definition of 'effective semantic diversity' used here, including the threshold for dsem.

Circularity Check

0 steps flagged · score 2.0 of 10

No formal circularity: the main accuracy and diversity results are measured on external benchmarks, and the self-generated RL supervision is a data-quality limitation, not a definitional equivalence.

full rationale

The paper's derivation chain is: synthesize MathV-DP from MultiMath-300K via DeepSeek-R1; supervised fine-tune on correct diverse trajectories; apply GRPO with rule-based accuracy, format, and discrimination/preference rewards; evaluate on MathVista, Math-V, and MMMU. The central accuracy claims are tested on held-out external benchmarks, so the reported gains (e.g., MathVista 68.2 to 70.4) are not by construction equal to the training signal. The diversity numbers are computed with a Sentence Transformer over the model's own generated responses on MathVista, not on the training labels, so they are also externally measured. The discrimination/preference rewards do use the same DeepSeek-R1-generated correct/incorrect labels as the SFT data, which is a legitimate concern about supervision provenance and label noise; however, the reward is a binary classification objective on those labels, and the paper does not present the reward value itself as the evidence for the claimed accuracy or diversity gains. The self-citations (e.g., Math-LLaVA [45], GalleryGPT [31]) appear in related work and are not used to justify the method's effectiveness. No equation reduces to its input by construction, and no fitted parameter is renamed as a prediction. The absence of a single-perspective control weakens the causal attribution to diversity, but that is an experimental-design limitation, not circularity. Overall, the paper is self-contained against external benchmarks, with only minor non-load-bearing self-citations and self-generated supervision caveats.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces a dataset and a trained model, but no new physical or conceptual entity such as a particle or force; these are artifacts, not entities in the sense of this ledger.

free parameters (3)
  • Semantic dissimilarity threshold in dsem = Not specified
    The effective semantic diversity metric treats two responses as dissimilar (distance 1) above an unstated threshold; this directly determines the reported diversity improvements in Table III.
  • Reward combination weights = Not specified
    The paper introduces three reward functions (accuracy, think-format, discrimination/preference) but does not state how they are weighted or combined in the GRPO objective; these weights affect the trained policy.
  • KL regularization coefficient beta = 0.04
    The KL coefficient in Eq. 6 is set to 0.04 by hand; while standard, it is a free choice that influences training stability and final performance.
assumptions (5)
  • domain assumption DeepSeek-R1 generates correctly labeled and genuinely diverse solutions from the given image captions, questions, and original solutions.
    The entire MathV-DP dataset and the RL reward labels rest on this; Section III-A and V-E provide only a limited human check.
  • domain assumption The formal dense captions in MultiMath-300K faithfully represent the visual content needed to solve each problem.
    The synthesis prompt conditions on these captions rather than the raw images; if captions omit key visual details, the generated solutions can be wrong; Section III-A.
  • domain assumption Sentence Transformer embeddings with a pairwise distance threshold capture the diversity of reasoning perspectives.
    The diversity metric in Eq. 7 uses dsem as a semantic distance function; no validation of this proxy is provided; Section IV-B.
  • standard math GRPO converges to a policy that appropriately balances the multiple reward signals.
    The paper relies on the GRPO objective from DeepSeekMath (ref [26]) and standard policy gradient theory.
  • standard math The answer extraction via regular expressions accurately matches model outputs to ground truth.
    Accuracy rewards are computed by regex-based answer matching; Section IV-A.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Mathematical Reasoning with Diverse Solving Perspective." pith.science (2026). https://pith.science/paper/MFQJGQMC

@misc{pith2026250702804,
  author       = {Pith},
  title        = {Pith review of: Multimodal Mathematical Reasoning with Diverse Solving Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFQJGQMC}},
  note         = {Machine review of arXiv:2507.02804}
}
read the original abstract

Recent progress in large-scale reinforcement learning (RL) has notably enhanced the reasoning capabilities of large language models (LLMs), especially in mathematical domains. However, current multimodal LLMs (MLLMs) for mathematical reasoning often rely on one-to-one image-text pairs and single-solution supervision, overlooking the diversity of valid reasoning perspectives and internal reflections. In this work, we introduce MathV-DP, a novel dataset that captures multiple diverse solution trajectories for each image-question pair, fostering richer reasoning supervision. We further propose Qwen-VL-DP, a model built upon Qwen-VL, fine-tuned with supervised learning and enhanced via group relative policy optimization (GRPO), a rule-based RL approach that integrates correctness discrimination and diversity-aware reward functions. Our method emphasizes learning from varied reasoning perspectives and distinguishing between correct yet distinct solutions. Extensive experiments on the MathVista's minitest and Math-V benchmarks demonstrate that Qwen-VL-DP significantly outperforms prior base MLLMs in both accuracy and generative diversity, highlighting the importance of incorporating diverse perspectives and reflective reasoning in multimodal mathematical reasoning.

Figures

Figures reproduced from arXiv: 2507.02804 by the authors.

Figure 1
Figure 1. An multimodal mathematical reasoning example with alternative [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall flowchart of the proposed multimodal question-solution data synthesis and post-training. Post-training consists of supervised fine-tuning and [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The prompt template used in our DeepSeek-R1 API for generating [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Accuracy of Qwen-VL model adopting different post-training strategies [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

75 extracted references · 33 canonical work pages

  1. [1]

    Self-consistency improves chain of thought reasoning in language models,

    X. Wang, J. Wei, D. Schuurmans, Q. V . Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou, “Self-consistency improves chain of thought reasoning in language models,” in The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  2. [2]

    Least-to-most prompting enables complex reasoning in large language models,

    D. Zhou et al., “Least-to-most prompting enables complex reasoning in large language models,” in The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  3. [3]

    Step 3.5 flash: Open frontier-level intelligence with 11b active parameters,

    A. Huang et al., “Step 3.5 flash: Open frontier-level intelligence with 11b active parameters,” arXiv preprint arXiv:2602.10604, 2026

    arXiv 2026

  4. [4]

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

    H. Luo, Q. Sun, C. Xu, P. Zhao, J. Lou, C. Tao, X. Geng, Q. Lin, S. Chen, and D. Zhang, “Wizardmath: Empowering mathematical reasoning for large language models via reinforced evol-instruct,” CoRR, vol. abs/2308.09583, 2023

    arXiv 2023

  5. [5]

    Mammoth: Building math generalist models through hybrid instruction tuning,

    X. Yue, X. Qu, G. Zhang, Y . Fu, W. Huang, H. Sun, Y . Su, and W. Chen, “Mammoth: Building math generalist models through hybrid instruction tuning,” CoRR, vol. abs/2309.05653, 2023

    arXiv 2023

  6. [6]

    Tora: A tool-integrated reasoning agent for mathematical problem solving,

    Z. Gou, Z. Shao, Y . Gong, Y . Shen, Y . Yang, M. Huang, N. Duan, and W. Chen, “Tora: A tool-integrated reasoning agent for mathematical problem solving,” CoRR, vol. abs/2309.17452, 2023

    arXiv 2023

  7. [7]

    Draft, sketch, and prove: Guiding formal theorem provers with informal proofs,

    A. Q. Jiang, S. Welleck, J. P. Zhou, T. Lacroix, J. Liu, W. Li, M. Jamnik, G. Lample, and Y . Wu, “Draft, sketch, and prove: Guiding formal theorem provers with informal proofs,” in The Eleventh International Conference on Learning Representations, 2023

    work page 2023

  8. [8]

    Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models,

    P. Lu, H. Bansal, T. Xia, J. Liu, C. Li, H. Hajishirzi, H. Cheng, K. Chang, M. Galley, and J. Gao, “Mathvista: Evaluating math reasoning in visual contexts with gpt-4v, bard, and other large multimodal models,” CoRR, vol. abs/2310.02255, 2023

    arXiv

Show all 75 references
  1. [9]

    Visual instruction tuning,

    H. Liu, C. Li, Q. Wu, and Y . J. Lee, “Visual instruction tuning,” in Advances in Neural Information Processing Systems, A. Oh, T. Nau- mann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., 2023

    2023

  2. [10]

    Qwen-vl: A frontier large vision-language model with versatile abilities,

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A frontier large vision-language model with versatile abilities,” CoRR, vol. abs/2308.12966, 2023

    2023 arXiv

  3. [11]

    UNK-VQA: A dataset and A probe into multi-modal large models’ abstention ability,

    Y . Guo, F. Jiao, Z. Shen, L. Nie, and M. S. Kankanhalli, “UNK-VQA: A dataset and A probe into multi-modal large models’ abstention ability,” CoRR, vol. abs/2310.10942, 2023

    2023 arXiv

  4. [12]

    OpenAI-b, “Gpt-4v,” https://openai.com/research/gpt-4v-system-card

  5. [13]

    Google, “Gemini,” https://gemini.google.com

  6. [14]

    Verbal reports as data

    K. A. Ericsson and H. A. Simon, “Verbal reports as data.” Psychological review, vol. 87, no. 3, p. 215, 1980

    1980

  7. [15]

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

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, B. Ichter, F. Xia, E. H. Chi, Q. V . Le, and D. Zhou, “Chain-of-thought prompting elicits reasoning in large language models,” in Advances in Neural Information Processing Systems, 2022

    2022

  8. [16]

    Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes,

    C.-Y . Hsieh, C.-L. Li, C.-K. Yeh, H. Nakhost, Y . Fujii, A. Ratner, R. Kr- ishna, C.-Y . Lee, and T. Pfister, “Distilling step-by-step! outperforming larger language models with less training data and smaller model sizes,” arXiv preprint arXiv:2305.02301, 2023

    2023 arXiv

  9. [17]

    Introducing openai o1,

    OpenAI-c, “Introducing openai o1,” https://openai.com/o1/

  10. [18]

    Instructblip: Towards general-purpose vision-language mod- els with instruction tuning,

    W. Dai, J. Li, D. Li, A. M. H. Tiong, J. Zhao, W. Wang, B. Li, P. N. Fung, and S. Hoi, “Instructblip: Towards general-purpose vision-language mod- els with instruction tuning,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  11. [19]

    R. S. Sutton, A. G. Barto et al., Reinforcement learning: An introduction. MIT press Cambridge, 1998, vol. 1, no. 1

    1998

  12. [20]

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

    D. Guo et al., “Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning,” arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  13. [21]

    Language models are unsupervised multitask learners,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, I. Sutskever et al., “Language models are unsupervised multitask learners,” OpenAI blog, vol. 1, no. 8, p. 9, 2019

    2019

  14. [22]

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

    L. Chen, L. Li, H. Zhao, Y . 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

  15. [23]

    R1-onevision: Advancing generalized multimodal reason- ing through cross-modal formalization,

    Y . Yanget al., “R1-onevision: Advancing generalized multimodal reason- ing through cross-modal formalization,” arXiv preprint arXiv:2503.10615, 2025

    2025 arXiv

  16. [24]

    Qwen2.5-vl technical report,

    S. Bai et al., “Qwen2.5-vl technical report,” arXiv preprint arXiv:2502.13923, 2025

    2025 arXiv

  17. [25]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,

    P. Wanget al., “Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution,” arXiv preprint arXiv:2409.12191, 2024

    2024 arXiv

  18. [26]

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

    Z. Shao et al., “Deepseekmath: Pushing the limits of mathematical reasoning in open language models,” arXiv preprint arXiv:2402.03300, 2024

    2024 arXiv

  19. [27]

    Measuring multimodal mathematical reasoning with math-vision dataset,

    K. Wang, J. Pan, W. Shi, Z. Lu, M. Zhan, and H. Li, “Measuring multimodal mathematical reasoning with math-vision dataset,” arXiv preprint arXiv:2402.14804, 2024

    2024 arXiv

  20. [28]

    Learning transferable visual models from natural language supervision,

    A. Radford et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning, 2021, pp. 8748–8763

    2021

  21. [29]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

    J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” in International conference on machine learning, 2022, pp. 12 888–12 900

    2022

  22. [30]

    Mm-forecast: A multimodal approach to temporal event forecasting with large language models,

    H. Li, Z. Yang, Y . Ma, Y . Bin, Y . Yang, and T.-S. Chua, “Mm-forecast: A multimodal approach to temporal event forecasting with large language models,” in ACM Multimedia 2024, 2024

    2024

  23. [31]

    Gallerygpt: Analyzing paintings with large multimodal models,

    Y . Bin, W. Shi, Y . Ding, Z. Hu, Z. Wang, Y . Yang, S.-K. Ng, and H. T. Shen, “Gallerygpt: Analyzing paintings with large multimodal models,” in ACM Multimedia 2024, 2024

    2024

  24. [32]

    Minigpt-4: Enhancing vision-language understanding with advanced large language models,

    D. Zhu, J. Chen, X. Shen, X. Li, and M. Elhoseiny, “Minigpt-4: Enhancing vision-language understanding with advanced large language models,” CoRR, vol. abs/2304.10592, 2023

    2023 arXiv

  25. [33]

    mplug-owl: Modularization empowers large language models with multimodality,

    Q. Ye et al., “mplug-owl: Modularization empowers large language models with multimodality,” arXiv preprint arXiv:2304.14178, 2023

    2023 arXiv

  26. [34]

    Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models,

    Z. Lin et al., “Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models,” arXiv preprint arXiv:2311.07575, 2023

    2023 arXiv

  27. [35]

    Minicpm: Unveiling the potential of small language models with scalable training strategies,

    S. Hu et al., “Minicpm: Unveiling the potential of small language models with scalable training strategies,” arXiv preprint arXiv:2404.06395, 2024

    2024 arXiv

  28. [36]

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

    W. Huang, B. Jia, Z. Zhai, S. Cao, Z. Ye, F. Zhao, Z. Xu, X. Tang, Y . Hu, and S. Lin, “Vision-r1: Incentivizing reasoning capability in multimodal large language models,” arXiv preprint arXiv:2503.06749, 2025

    2025 arXiv

  29. [37]

    Adapting multimodal large language models for video question answering by capturing question-critical and coherent moments,

    H. Wang, C. Lai, and W. Ge, “Adapting multimodal large language models for video question answering by capturing question-critical and coherent moments,” IEEE Transactions on Multimedia, 2025

    2025

  30. [38]

    Mmtot: Multi- modal token-of-thought reasoning for large models,

    N. Xu, Z. Lu, H. Tian, B. Zheng, J. Cao, and A.-A. Liu, “Mmtot: Multi- modal token-of-thought reasoning for large models,” IEEE Transactions on Multimedia, 2026

    2026

  31. [39]

    Multimodal chain-of-thought reasoning in language models,

    Z. Zhang, A. Zhang, M. Li, H. Zhao, G. Karypis, and A. Smola, “Multimodal chain-of-thought reasoning in language models,” arXiv preprint arXiv:2302.00923, 2023

    2023 arXiv

  32. [40]

    T- sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering,

    L. Wang, Y . Hu, J. He, X. Xu, N. Liu, H. Liu, and H. T. Shen, “T- sciq: Teaching multimodal chain-of-thought reasoning via large language model signals for science question answering,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 38, 2024, pp. 19 162– 19 170

    2024

  33. [41]

    Chain-of-thought prompt distillation for multi- modal named entity and multimodal relation extraction,

    F. Chen and Y . Feng, “Chain-of-thought prompt distillation for multi- modal named entity and multimodal relation extraction,” arXiv preprint arXiv:2306.14122, 2023

    2023 arXiv

  34. [42]

    Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models,

    L. Li, Y . Wang, R. Xu, P. Wang, X. Feng, L. Kong, and Q. Liu, “Multimodal ArXiv: A dataset for improving scientific comprehension of large vision-language models,” in Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 2024, pp. 14 369–14 387

    2024

  35. [43]

    Visual program distillation: Distilling tools and programmatic reasoning into vision-language models,

    Y . Hu, O. Stretcu, C. Lu, K. Viswanathan, K. Hata, E. Luo, R. Krishna, and A. Fuxman, “Visual program distillation: Distilling tools and programmatic reasoning into vision-language models,” CoRR, vol. abs/2312.03052, 2023

    2023 arXiv

  36. [44]

    Ddcot: Duty- distinct chain-of-thought prompting for multimodal reasoning in language models,

    G. Zheng, B. Yang, J. Tang, H.-Y . Zhou, and S. Yang, “Ddcot: Duty- distinct chain-of-thought prompting for multimodal reasoning in language models,” Advances in Neural Information Processing Systems, vol. 36, pp. 5168–5191, 2023

    2023

  37. [45]

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

    W. Shi, Z. Hu, Y . Bin, J. Liu, Y . Yang, S.-K. Ng, L. Bing, and R. K.-W. Lee, “Math-llava: Bootstrapping mathematical reasoning for multimodal large language models,” arXiv preprint arXiv:2406.17294, 2024

    2024 arXiv

  38. [46]

    MMMU: A massive multi-discipline multimodal un- derstanding and reasoning benchmark for expert AGI,

    X. Yue et al., “MMMU: A massive multi-discipline multimodal un- derstanding and reasoning benchmark for expert AGI,” CoRR, vol. abs/2311.16502, 2023

    2023 arXiv

  39. [47]

    Reinforcement learning: A survey, journal of artificial intelligence research 4,

    M. Littman and A. Moore, “Reinforcement learning: A survey, journal of artificial intelligence research 4,” 1996

    1996

  40. [48]

    Language models are few-shot learners,

    T. Brown et al., “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

    1901

  41. [49]

    Training a helpful and harmless assistant with reinforcement learning from human feedback,

    Y . Baiet al., “Training a helpful and harmless assistant with reinforcement learning from human feedback,” arXiv preprint arXiv:2204.05862, 2022

    2022 arXiv

  42. [50]

    Prox- imal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Prox- imal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  43. [51]

    Direct preference optimization: Your language model is secretly a reward model,

    R. Rafailov, A. Sharma, E. Mitchell, C. D. Manning, S. Ermon, and C. Finn, “Direct preference optimization: Your language model is secretly a reward model,” Advances in Neural Information Processing Systems, vol. 36, pp. 53 728–53 741, 2023. JOURNAL OF LATEX CLASS FILES, VOL. ...

    2023

  44. [52]

    Kimi k1. 5: Scaling reinforcement learning with llms,

    K. Team et al., “Kimi k1. 5: Scaling reinforcement learning with llms,” arXiv preprint arXiv:2501.12599, 2025

    2025 arXiv

  45. [53]

    Ursa: Understanding and verifying chain-of-thought reasoning in multimodal mathematics,

    R. Luo, Z. Zheng, Y . Wang, Y . Yu, X. Ni, Z. Lin, J. Zeng, and Y . Yang, “Ursa: Understanding and verifying chain-of-thought reasoning in multimodal mathematics,” arXiv e-prints, pp. arXiv–2501, 2025

    2025

  46. [54]

    More than one teacher: Adaptive multi-guidance policy optimization for diverse exploration,

    X. Yuan, Y . Ding, Y . Bin, W. Shao, J. Cai, J. Song, Y . Yang, and H. T. Shen, “More than one teacher: Adaptive multi-guidance policy optimization for diverse exploration,” arXiv preprint arXiv:2510.02227, 2025

    2025

  47. [55]

    Sophiavl-r1: Reinforcing mllms reasoning with thinking reward,

    K. Fan, K. Feng, H. Lyu, D. Zhou, and X. Yue, “Sophiavl-r1: Reinforcing mllms reasoning with thinking reward,” arXiv preprint arXiv:2505.17018, 2025

    2025

  48. [56]

    From sight to insight: Improving visual reasoning capabilities of multimodal models via reinforcement learning,

    O. Sharif, E. Hossain, and P. Ng, “From sight to insight: Improving visual reasoning capabilities of multimodal models via reinforcement learning,” arXiv preprint arXiv:2601.00215, 2026

    2026 arXiv

  49. [57]

    Rest- mcts*: Llm self-training via process reward guided tree search,

    D. Zhang, S. Zhoubian, Z. Hu, Y . Yue, Y . Dong, and J. Tang, “Rest- mcts*: Llm self-training via process reward guided tree search,”Advances in Neural Information Processing Systems, vol. 37, pp. 64 735–64 772, 2024

    2024

  50. [58]

    Multimath: Bridging visual and mathematical reasoning for large language models,

    S. Peng, D. Fu, L. Gao, X. Zhong, H. Fu, and Z. Tang, “Multimath: Bridging visual and mathematical reasoning for large language models,” arXiv preprint arXiv:2409.00147, 2024

    2024 arXiv

  51. [59]

    G-llava: Solving geometric problem with multi-modal large language model,

    J. Gao et al., “G-llava: Solving geometric problem with multi-modal large language model,” arXiv preprint arXiv:2312.11370, 2023

    2023 arXiv

  52. [60]

    Gemini: a family of highly capable multimodal models,

    G. Team et al., “Gemini: a family of highly capable multimodal models,” arXiv preprint arXiv:2312.11805, 2023

    2023 arXiv

  53. [61]

    The claude 3 model family: Opus, sonnet, haiku,

    A. Anthropic, “The claude 3 model family: Opus, sonnet, haiku,” Claude-3 Model Card, 2024

    2024

  54. [62]

    Chatgpt,

    OpenAI-a, “Chatgpt,” https://chat.openai.com

  55. [63]

    The llama 4 herd: The beginning of a new era of natively multimodal ai innovation,

    Meta-AI, “The llama 4 herd: The beginning of a new era of natively multimodal ai innovation,” https://ai.meta.com/blog/ llama-4-multimodal-intelligence/

  56. [64]

    Obelics: An open web-scale filtered dataset of interleaved image-text documents,

    H. Laurençon et al., “Obelics: An open web-scale filtered dataset of interleaved image-text documents,” Advances in Neural Information Processing Systems, vol. 36, 2024

    2024

  57. [65]

    Llama-adapter v2: Parameter-efficient visual instruction model,

    P. Gao et al., “Llama-adapter v2: Parameter-efficient visual instruction model,” arXiv preprint arXiv:2304.15010, 2023

    2023 arXiv

  58. [66]

    Llavar: Enhanced visual instruction tuning for text-rich image under- standing,

    Y . Zhang, R. Zhang, J. Gu, Y . Zhou, N. Lipka, D. Yang, and T. Sun, “Llavar: Enhanced visual instruction tuning for text-rich image under- standing,” arXiv preprint arXiv:2306.17107, 2023

    2023 arXiv

  59. [67]

    Improved baselines with visual instruction tuning,

    H. Liu, C. Li, Y . Li, and Y . J. Lee, “Improved baselines with visual instruction tuning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 26 296–26 306

    2024

  60. [68]

    Large multi-modal models for strong performance and efficient deployment

    OpenBMB, “Large multi-modal models for strong performance and efficient deployment.” 2024, https://github.com/OpenBMB/OmniLMM

    2024

  61. [69]

    Llava-next: Improved reasoning, ocr, and world knowledge,

    H. Liu, C. Li, Y . Li, B. Li, Y . Zhang, S. Shen, and Y . J. Lee, “Llava-next: Improved reasoning, ocr, and world knowledge,” 2024. [Online]. Available: https://llava-vl.github.io/blog/2024-01-30-llava-next/

    2024

  62. [70]

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

    W. Zhuang, X. Huang, X. Zhang, and J. Zeng, “Math-puma: Progressive upward multimodal alignment to enhance mathematical reasoning,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 24, 2025, pp. 26 183–26 191

    2025

  63. [71]

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

    H. Yao 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

  64. [72]

    Evaluating the diversity and quality of llm generated content,

    A. Shypula, S. Li, B. Zhang, V . Padmakumar, K. Yin, and O. Bastani, “Evaluating the diversity and quality of llm generated content,” arXiv preprint arXiv:2504.12522, 2025

    2025

  65. [73]

    Sentence-bert: Sentence embeddings using siamese bert-networks,

    N. Reimers and I. Gurevych, “Sentence-bert: Sentence embeddings using siamese bert-networks,” in Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 11 2019

    2019

  66. [74]

    On llms-driven synthetic data generation, curation, and evaluation: A survey,

    L. Long, R. Wang, R. Xiao, J. Zhao, X. Ding, G. Chen, and H. Wang, “On llms-driven synthetic data generation, curation, and evaluation: A survey,” arXiv preprint arXiv:2406.15126, 2024. VII. BIOGRAPHYSECTION Wenhao Shireceived the M.Sc. degree in computer science and B.Sc. deg...

    2024 arXiv

  67. [1991]

    His research interests include artificial intelligence, data science, machine learning, multimodal intelligence, bioinformatics, and smart city analytics

    He is currently a Professor of Practice with the Department of Computer Science, School of Computing, National University of Singapore, Singapore, and the Director of Translational Research at the Institute of Data Science, NUS. His research interests include artificial intell...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.