Pith. sign in

REVIEW 5 major objections 4 minor 4 cited by

GM-PRM: A Generative Multimodal Process Reward Model for Multimodal Mathematical Reasoning

T0 review · 5 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read This paper argues that a process reward model for multimodal mathematics should generate corrections, not just scores, and that correcting the first wrong step at test time improves six different multimodal language models on five benchmark

desk verdict GM-PRM is a genuine extension of generative PRMs to multimodal math with a corrective BoN loop, but the SOTA claim rests on an uncontaminated eval set and direct comparisons that the paper doesn't provide. read the letter →

arxiv 2508.04088 v2 pith:C3EE7H3E submitted 2025-08-06 cs.CL

classification cs.CL
keywords multimodalmathreasoningprocessrewardmodelgenerativetest-timescalingBest-of-Nstep-levelcorrectionvisualalignmentchain-of-thoughtverification
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

GM-PRM turns the process reward model from a passive verifier into an active corrector for multimodal math. The paper claims that a 7B model trained on 19,614 plane-geometry and function problems can, for each reasoning step, state the step's intent, check its visual alignment, judge its logic, and, when it finds the first wrong step, generate a corrected version of that step. At test time, Refined Best-of-N uses those corrections to reseed the policy model's next batch of solutions, producing a more diverse and accurate pool. Across five benchmarks and six different MLLMs, the paper reports average accuracy gains of 2.9 to 5.9 points and a larger Pass@8 than ordinary Best-of-N. If correct, this shows that test-time compute can be spent on repairing flawed reasoning rather than only ranking finished answers, with a surprisingly small training set.

What carries the argument

The load-bearing object is GM-PRM, a generative multimodal process reward model built from a 7B vision-language model and fine-tuned on teacher-generated critiques that were filtered by agreement with Monte Carlo completion labels. It maps (question, image, reasoning steps) to per-step textual critiques and judgments in four dimensions: step intent, image alignment, reasoning logic, and step refinement, plus a refined replacement of the first erroneous step. The companion mechanism is Refined-BoN, which feeds those corrections back into the policy model instead of passively discarding flawed solutions; this is what lets the reward model actively change the solution pool rather than just rank

What would settle it

Build a held-out set of geometry and function problems whose solutions have human-annotated first errors, especially subtle image-alignment errors such as a misread angle or a misidentified parallel line. Count how often GM-PRM names the exact human-annotated first error and whether the policy model, continuing from GM-PRM's replacement step, reaches the correct answer. If corrected continuations on this set do not beat ordinary Best-of-N on the same problems, the reported gains come from reranking, not from the correction mechanism.

Watch

Extended reading notes

Core claim

The central claim is that the reason existing multimodal PRMs plateau is that they can say 'wrong' but not 'here is the fix.' GM-PRM is trained to do both: per step, it outputs a step-intent analysis, an image-alignment analysis, a reasoning-logic analysis, a binary-style judgment, and, only for the first erroneous step, a concrete refined step. The paper then wraps this critic in Refined-BoN: the first N/2 solutions are critiqued; for each flawed one, the critic stops at the first error, replaces it, and asks the policy model to continue from the validated prefix; the remaining solutions are generated from those repaired trajectories. Final selection averages the correctness-token probabili

Load-bearing premise

The whole method rests on the training data teaching GM-PRM to find the first genuinely wrong step and replace it with a truly correct one; if the filtered teacher critiques mostly capture easy or formulaic slips, the corrections will not fix the subtle visual-alignment and logic errors that decide benchmark scores.

Editorial extensions

If this is right

  • Test-time compute can be spent on repairing reasoning, not only on sampling and ranking: a flawed solution is edited at its first bad step and the policy model continues from the repaired prefix.
  • A single 7B critic transfers across policy models: the same GM-PRM improves six MLLMs without per-model retraining, and weaker baselines gain the most.
  • Data efficiency is high: about 20K curated critique/correction samples suffice, shifting the bottleneck from dataset size to the quality of teacher critiques.
  • Averaging per-step correctness-token probabilities is a better solution-selection rule than taking the max or min step score.
  • Larger solution budgets amplify the advantage: as N grows from 1 to 8, Refined-BoN gains accuracy faster than self-consistency.

Reading between the lines

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

  • The paper does not test repeated refinement: after a repaired trajectory is generated, GM-PRM could critique it again; iterating the loop would either push accuracy further or saturate at the teacher's skill ceiling.
  • The training set covers only plane geometry and functions, yet gains appear on other categories; this hints the method may transfer to other visually grounded symbolic tasks such as charts, tables, or scientific diagrams, though the paper's evidence there is limited.
  • The correction ceiling is set by the teacher that wrote the training critiques: a stronger or weaker teacher should shift the gains accordingly, making teacher substitution the cleanest way to measure how much of the effect is due to correction rather than reranking.
  • The natural-language critiques are an interpretability byproduct: a step-level 'why this is wrong' explanation could be used for human debugging or targeted fine-tuning, which the paper does not count in its accuracy numbers.
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

5 major / 4 minor

Summary. GM-PRM replaces scalar process-reward scores with generative, multi-aspect critiques (step intent, image alignment, reasoning logic) and a corrected version of the first erroneous step. The model is produced by SFT of Qwen2.5-VL-7B on 19,614 samples drawn from VisualPRM400K, annotated by GPT-4o and filtered by agreement with Monte Carlo labels. At test time, Refined-BoN generates N/2 initial solutions, uses GM-PRM to correct flawed prefixes, samples additional solutions conditioned on those corrections, and selects among the N candidates by averaging step-level token probabilities. The paper reports consistent accuracy gains for six policy MLLMs across MathVista, MathVision, MathVerse Vision-Only, DynaMath Plane-Geometry, and WeMath, with average improvements of +2.9 to +5.9 points.

Significance. The core idea is useful and forward-looking: a generative multimodal PRM that corrects rather than only ranks steps could make test-time scaling more effective and more interpretable. The paper deserves credit for evaluating across six policy models and five benchmarks, for ablating the number of candidates and score aggregation, and for planning to release code. However, the current evidence does not establish the 'state-of-the-art' claim, and the possible overlap between the VisualPRM400K training subset and the evaluation benchmarks is an unresolved threat to the central result. If the gains survive decontamination and are compared fairly with existing multimodal PRMs, this would be a solid contribution.

major comments (5)
  1. [§3.2, §4.1, A.3] The central claim is threatened by possible train/eval overlap. The training set is a 19,614-sample subset of VisualPRM400K (A.3) selected for plane-geometry and function problems (3.2), while evaluation uses MathVista, MathVision, MathVerse, DynaMath, and WeMath (4.1, Table 3). The paper does not report whether any VisualPRM400K instances, or their images, appear in these test splits. Because GM-PRM is trained to output corrected steps, memorized benchmark solutions would inflate Refined-BoN gains. Please report a decontamination analysis (exact and near-duplicate matching of question text and images, per benchmark) and rerun the main tables with any overlapping samples removed.
  2. [§4.2, Table 1] The 'state-of-the-art' wording in the Abstract and Conclusion is not supported by the experimental design. Table 1 compares each policy model only to its own unguided sampling; it contains no comparison against existing multimodal PRMs (VisualPRM, URSA, Athena-PRM, MM-PRM, DreamPRM, all listed in §2), nor against a scalar PRM baseline under the same BoN budget. Add such comparisons, including the gain relative to the best existing PRM under matched sampling conditions, before claiming SOTA.
  3. [§3.3] Refined-BoN is under-specified. The text says N/2 initial solutions are generated, then the 'subsequent N/2' are produced by feeding corrected steps plus 'all previously validated correct steps' back to the policy model, but it does not give an algorithm: how many continuations are sampled per corrected prefix, whether the corrected prefix is the only new condition, how a fully correct solution is used to generate a new solution, and how solution-level scores are computed when some steps are unscored. The final-selection paragraph also conflates 'all steps judged correct' with scoring all steps, while §4.3 later states that solutions with incorrect steps are often not scored for all steps. Provide pseudocode and exact prompt/selection rules.
  4. [§3.2, Eq. (5)] The correction target is generated by GPT-4o and filtered only by agreement between GPT-4o judgments and MC labels; the corrected step itself is not verified. A wrong GPT-4o refinement can still pass the filter if the step judgment is 'incorrect'. Validate corrected steps (e.g., MC completion from the corrected prefix, or human spot checks) and report agreement rates. Otherwise the model may learn superficially plausible but wrong corrections, which is a central component of the claimed gains.
  5. [§4.2, Table 1] No variance or significance information is given: all numbers are single runs, and several improvements are small (+1.6, +1.3, +1.8). Report multiple seeds or bootstrap confidence intervals, together with paired significance tests, for both the aggregate and per-benchmark gains. Without this, the claim of 'consistent gains' is not statistically grounded.
minor comments (4)
  1. [Table 1 caption] The caption says 'All values are reported after rounding to three decimal places,' but the table shows one decimal place. Please correct.
  2. [Eq. (8)] 'fcomleter' is a typo for 'fcompleter'.
  3. [Table 3 caption] 'Benchamrks' is a typo for 'Benchmarks'.
  4. [§4.3, Figures 4 and 5] The figures would be much more informative with error bars or multiple-run markers; as presented, the differences could be within run-to-run variation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GM-PRM is trained on external annotations and evaluated on benchmark accuracy; self-citations and self-scoring do not reduce the central claim to its inputs.

full rationale

The paper's derivation chain is empirical rather than definitional. GM-PRM is trained on a subset of VisualPRM400K; supervision comes from GPT-4o-generated critiques plus Monte-Carlo hard labels (Eqs. 4-9), and the training/filtering procedure does not use the evaluation benchmarks' ground-truth answers. The central claim—that GM-PRM + Refined-BoN improves policy-model accuracy—is measured against external benchmark answers (Table 1), so it is not a restatement of the training signal. The Refined-BoN selection uses GM-PRM's own step scores, and the paper's self-citations (Yan et al. 2024a,b,c, 2025a,b; Zheng et al. 2024b) appear only in introductory and related-work context and do not carry the method's validity. A residual risk of train/eval overlap with VisualPRM400K is a data-contamination concern, not a circularity, and cannot be established from the text; per the review rules it does not raise the circularity score.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim is an empirical accuracy improvement, so the ledger records the hand-chosen test-time constants and the trust assumptions about the GPT-4o/MC supervision. No new physical entities are introduced; GM-PRM is a model, not a postulated entity.

free parameters (3)
  • Number of candidates N = 8
    Set by hand in Refined-BoN; Figure 4 shows accuracy rises with N, so the reported gains are evaluated at this specific budget.
  • Sampling hyperparameters = temperature=0.7, top-p=0.9
    Used for policy model generation in all experiments; not tuned per benchmark.
  • Data filtering agreement criterion = exact agreement between GPT-4o judgment and hard MC label
    Section 3.2 keeps samples where both judges agree; this choice shapes the training distribution and correction quality.
assumptions (4)
  • domain assumption GPT-4o critiques and corrections are accurate enough to serve as supervision for GM-PRM
    Section 3.2 constructs the entire training set from GPT-4o outputs; no human quality check is reported.
  • domain assumption Hard Monte Carlo estimation labels are reliable indicators of step correctness
    Section 3.2 uses Equation 9 hard estimation to filter; MC labels can be noisy when completers are weak.
  • domain assumption Benchmark ground truths are correct and independent of the training data
    Section 4.1 uses five public benchmarks; if train/eval overlap exists it is not discussed.
  • domain assumption The corrected prefix fed back to the policy model yields a valid continuation
    Section 3.3 assumes the policy model can continue from a corrected step without degenerating; no analysis of continuation failures is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of GM-PRM: A Generative Multimodal Process Reward Model for Multimodal Mathematical Reasoning." pith.science (2026). https://pith.science/paper/C3EE7H3E

@misc{pith2026250804088,
  author       = {Pith},
  title        = {Pith review of: GM-PRM: A Generative Multimodal Process Reward Model for Multimodal Mathematical Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C3EE7H3E}},
  note         = {Machine review of arXiv:2508.04088}
}
read the original abstract

Multimodal Large Language Models (MLLMs) demonstrate remarkable capabilities but often struggle with complex, multi-step mathematical reasoning, where minor errors in visual perception or logical deduction can lead to complete failure. While Process Reward Models (PRMs) offer step-by-step supervision, existing multimodal PRMs are limited to being binary verifiers that can identify but not correct errors, offering little explanatory power. To address these deficiencies, we introduce the Generative Multimodal Process Reward Model (GM-PRM), a novel paradigm that transforms the PRM from a passive judge into an active reasoning collaborator. Instead of a simple scalar score, GM-PRM provides a fine-grained, interpretable analysis of each reasoning step, evaluating its step intent, visual alignment, and logical soundness. More critically, GM-PRM is trained to generate a corrected version of the first erroneous step it identifies. This unique corrective capability enables our new test-time inference strategy, Refined Best-of-N (Refined-BoN). This framework actively enhances solution quality by using the PRM's generated correction to guide the policy model toward a more promising reasoning trajectory, thereby improving the diversity and correctness of the solution pool. We demonstrate that GM-PRM achieves state-of-the-art results on multiple multimodal math benchmarks, significantly boosting policy model performance with remarkable data efficiency, requiring only a 20K-sample training dataset. Our code will be released upon acceptance.

Figures

Figures reproduced from arXiv: 2508.04088 by the authors.

Figure 1
Figure 1. Comparison among ORM (a), PRM (b), and our [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The illustration of a representative example [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Improvements of the average percentage accuracy [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: The results of changing the value of N in the Refined-BoN process on the WeMath across different policies. As N [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Average percentage accuracy (%) of MLLMs via [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Prompt for policy models to generate reasoning and [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]
Figure 7
Figure 7. Figure 7: Prompt for policy models to generate reasoning and [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 8
Figure 8. Figure 8: Prompt for GPT-4o to construct training dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png]
Figure 9
Figure 9. Figure 9: The illustration of a representative example [PITH_FULL_IMAGE:figures/full_fig_p013_9.png]
Figure 10
Figure 10. Figure 10: The illustration of a representative example [PITH_FULL_IMAGE:figures/full_fig_p014_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 4 Pith papers

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

  1. Test-Time Scaling for Small VLMs on Multilingual Visual MCQ

    cs.CL 2026-07 accept novelty 5.5 of 10

    On EXAMS-V, token budget and parseability dominate chain count, PRM-guided search, and selectors for small VLMs; the policy model itself yields the largest gain.

  2. CogniRoute: Learning to Route Social Evidence in Omni-Modal Models

    cs.CV 2026-06 unverdicted novelty 5.0 of 10

    CogniRoute adds a cognitive schema and route-aware RL to an omni-modal MoE, reaching 59.38% accuracy on a new 118K-example social video QA benchmark and beating prior baselines by 15-27 points.

  3. PaLMR: Towards Faithful Visual Reasoning via Multimodal Process Alignment

    cs.CV 2026-02 conditional novelty 5.0 of 10

    By gating the reinforcement-learning reward on a binary visual-fidelity score, PaLMR lifts Qwen2.5-VL-7B to 70.9 on HallusionBench and 59.3 on MMMU, surpassing its GRPO baseline.

  4. Toward Native Multimodal Modeling: A Roadmap

    cs.CV 2026-05 unverdicted novelty 3.0 of 10

    A roadmap that defines architectural nativity for multimodal models and categorizes them into Multi-to-Text, Multi-to-Target, and Multi-to-Multi types while outlining an industrial pipeline toward unified transformer-...

Reference graph

Works this paper leans on

56 extracted references · 9 canonical work pages · cited by 4 Pith papers

  1. [1]

    Ahn, J.; Verma, R.; Lou, R.; Liu, D.; Zhang, R.; and Yin, W. 2024. Large language models for mathematical reasoning: Progresses and challenges. arXiv preprint arXiv:2402.00157

  2. [2]

    Bai, J.; Bai, S.; Chu, Y.; Cui, Z.; Dang, K.; Deng, X.; Fan, Y.; Ge, W.; Han, Y.; Huang, F.; Hui, B.; Ji, L.; Li, M.; Lin, J.; Lin, R.; Liu, D.; Liu, G.; Lu, C.; Lu, K.; Ma, J.; Men, R.; Ren, X.; Ren, X.; et al. 2023. Qwen Technical Report. arXiv:2309.16609

  3. [3]

    Bai, S.; Chen, K.; Liu, X.; Wang, J.; Ge, W.; Song, S.; Dang, K.; Wang, P.; Wang, S.; Tang, J.; Zhong, H.; Zhu, Y.; Yang, M.; Li, Z.; Wan, J.; Wang, P.; Ding, W.; Fu, Z.; Xu, Y.; Ye, J.; Zhang, X.; Xie, T.; et al. 2025. Qwen2.5-VL Technical Report. arXiv:2502.13923

  4. [4]

    Caffagni, D.; Cocchi, F.; Barsellotti, L.; Moratelli, N.; Sarto, S.; Baraldi, L.; Cornia, M.; and Cucchiara, R. 2024. The revolution of multimodal large language models: a survey. arXiv preprint arXiv:2402.12451

  5. [5]

    A.; and Xie, P

    Cao, Q.; Wang, R.; Zhang, R.; Somayajula, S. A.; and Xie, P. 2025. DreamPRM: Domain-Reweighted Process Reward Model for Multimodal Reasoning. arXiv preprint arXiv:2505.20241

  6. [6]

    Chen, Z.; Wu, J.; Wang, W.; Su, W.; Chen, G.; Xing, S.; Zhong, M.; Zhang, Q.; Zhu, X.; Lu, L.; Li, B.; Luo, P.; Lu, T.; Qiao, Y.; and Dai, J. 2024. InternVL: Scaling up Vision Foundation Models and Aligning for Generic Visual-Linguistic Tasks. arXiv:2312.14238

  7. [7]

    Du, L.; Meng, F.; Liu, Z.; Zhou, Z.; Luo, P.; Zhang, Q.; and Shao, W. 2025. MM-PRM: Enhancing Multimodal Mathematical Reasoning with Scalable Step-Level Supervision. arXiv preprint arXiv:2505.13427

  8. [8]

    Gao, J.; Xu, S.; Ye, W.; Liu, W.; He, C.; Fu, W.; Mei, Z.; Wang, G.; and Wu, Y. 2024. On designing effective rl reward at training time for llm reasoning. arXiv preprint arXiv:2410.15115

Show all 56 references
  1. [9]

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; Yang, A.; Fan, A.; Goyal, A.; Hartshorn, A.; Yang, A.; Mitra, A.; Sravankumar, A.; Korenev, A.; et al. 2024. The Llama 3 Herd of Models. arXiv:2407.21783

  2. [10]

    Hu, P.; Zhang, Z.; Chang, Q.; Liu, S.; Ma, J.; Du, J.; Zhang, J.; Liu, Q.; Gao, J.; Ma, F.; et al. 2025. PRM-BAS: Enhancing Multimodal Reasoning through PRM-guided Beam Annealing Search. arXiv preprint arXiv:2504.10222

  3. [11]

    Huo, J.; Yan, Y.; Hu, B.; Yue, Y.; and Hu, X. 2024. Mmneuron: Discovering neuron-level domain-specific interpretation in multimodal large language model. arXiv preprint arXiv:2406.11193

  4. [12]

    Khalifa, M.; Agarwal, R.; Logeswaran, L.; Kim, J.; Peng, H.; Lee, M.; Lee, H.; and Wang, L. 2025. Process Reward Models That Think. arXiv:2504.16828

  5. [13]

    Y.; Chandu, K.; Dziri, N.; Kumar, S.; Zick, T.; Choi, Y.; et al

    Lambert, N.; Pyatkin, V.; Morrison, J.; Miranda, L.; Lin, B. Y.; Chandu, K.; Dziri, N.; Kumar, S.; Zick, T.; Choi, Y.; et al. 2024. Rewardbench: Evaluating reward models for language modeling. arXiv preprint arXiv:2403.13787

  6. [14]

    Li, Q.; Chen, Z.; Wang, W.; Wang, W.; Ye, S.; Jin, Z.; Chen, G.; He, Y.; Gao, Z.; Cui, E.; Yu, J.; Tian, H.; Zhou, J.; Xu, C.; Wang, B.; Wei, X.; Li, W.; Zhang, W.; Zhang, B.; Cai, P.; Wen, L.; et al. 2024. OmniCorpus: A Unified Multimodal Corpus of 10 Billion-Level Images Int...

  7. [15]

    Li, W.; and Li, Y. 2024. Process reward model with q-value rankings. arXiv preprint arXiv:2410.11287

  8. [16]

    Li, X.; Yu, H.; Zhang, X.; Huang, Z.; He, S.; Liu, K.; Zhao, J.; Huang, F.; and Li, Y. 2025 a . Socratic-PRMBench: Benchmarking Process Reward Models with Systematic Reasoning Patterns. arXiv preprint arXiv:2505.23474

  9. [17]

    Li, Z.; Wen, X.; Lou, J.; Ji, Y.; Lu, Y.; Han, X.; Zhang, D.; and Sun, L. 2025 b . The devil is in the details: Tackling unimodal spurious correlations for generalizable multimodal reward models. arXiv preprint arXiv:2503.03122

  10. [18]

    Lightman, H.; Kosaraju, V.; Burda, Y.; Edwards, H.; Baker, B.; Lee, T.; Leike, J.; Schulman, J.; Sutskever, I.; and Cobbe, K. 2023. Let's verify step by step. In The Twelfth International Conference on Learning Representations

  11. [19]

    Liu, W.; Li, J.; Zhang, X.; Zhou, F.; Cheng, Y.; and He, J. 2024 a . Diving into Self-Evolving Training for Multimodal Reasoning. arXiv:2412.17451

  12. [20]

    Liu, Y.; Cao, Y.; Gao, Z.; Wang, W.; Chen, Z.; Wang, W.; Tian, H.; Lu, L.; Zhu, X.; Lu, T.; Qiao, Y.; and Dai, J. 2024 b . MMInstruct: a high-quality multi-modal instruction tuning dataset with extensive diversity. Science China Information Sciences, 67(12)

  13. [21]

    Liu, Z.; He, Y.; Wang, W.; Wang, W.; Wang, Y.; Chen, S.; Zhang, Q.; Lai, Z.; Yang, Y.; Li, Q.; Yu, J.; Li, K.; Chen, Z.; Yang, X.; Zhu, X.; Wang, Y.; Wang, L.; Luo, P.; Dai, J.; and Qiao, Y. 2023. InternGPT: Solving Vision-Centric Tasks by Interacting with ChatGPT Beyond Langu...

  14. [22]

    Lu, P.; Bansal, H.; Xia, T.; Liu, J.; Li, C.; Hajishirzi, H.; Cheng, H.; Chang, K.-W.; Galley, M.; and Gao, J. 2023. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255

  15. [23]

    Luo, L.; Liu, Y.; Liu, R.; Phatale, S.; Guo, M.; Lara, H.; Li, Y.; Shu, L.; Zhu, Y.; Meng, L.; Sun, J.; and Rastogi, A. 2024. Improve Mathematical Reasoning in Language Models by Automated Process Supervision. arXiv:2406.06592

  16. [24]

    Luo, R.; Zheng, Z.; Wang, Y.; Ni, X.; Lin, Z.; Jiang, S.; Yu, Y.; Shi, C.; Chu, R.; Zeng, J.; and Yang, Y. 2025. URSA: Understanding and Verifying Chain-of-thought Reasoning in Multimodal Mathematics. arXiv:2501.04686

  17. [25]

    Miao, B.; Wu, Y.; Gao, M.; Yu, Q.; Bu, W.; Zhang, W.; Li, Y.; Tang, S.; Chua, T.-S.; and Li, J. 2025. Boosting Virtual Agent Learning and Reasoning: A Step-Wise, Multi-Dimensional, and Generalist Reward Model with Benchmark. arXiv preprint arXiv:2503.18665

  18. [26]

    Qiao, R.; Tan, Q.; Dong, G.; Wu, M.; Sun, C.; Song, X.; GongQue, Z.; Lei, S.; Wei, Z.; Zhang, M.; et al. 2024. We-math: Does your large multimodal model achieve human-like mathematical reasoning? arXiv preprint arXiv:2407.01284

  19. [27]

    Setlur, A.; Nagpal, C.; Fisch, A.; Geng, X.; Eisenstein, J.; Agarwal, R.; Agarwal, A.; Berant, J.; and Kumar, A. 2024. Rewarding progress: Scaling automated process verifiers for llm reasoning. arXiv preprint arXiv:2410.08146

  20. [28]

    She, S.; Liu, J.; Liu, Y.; Chen, J.; Huang, X.; and Huang, S. 2025. R-PRM: Reasoning-Driven Process Reward Modeling. arXiv:2503.21295

  21. [29]

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

  22. [30]

    Song, M.; Su, Z.; Qu, X.; Zhou, J.; and Cheng, Y. 2025. PRMBench: A fine-grained and challenging benchmark for process-level reward models. arXiv preprint arXiv:2501.03124

  23. [31]

    Tian, C.; Zhu, X.; Xiong, Y.; Wang, W.; Chen, Z.; Wang, W.; Chen, Y.; Lu, L.; Lu, T.; Zhou, J.; Li, H.; Qiao, Y.; and Dai, J. 2024. MM-Interleaved: Interleaved Image-Text Generative Modeling via Multi-modal Feature Synchronizer. arXiv:2401.10208

  24. [32]

    Touvron, H.; Lavril, T.; Izacard, G.; Martinet, X.; Lachaux, M.-A.; Lacroix, T.; Rozière, B.; Goyal, N.; Hambro, E.; Azhar, F.; Rodriguez, A.; Joulin, A.; Grave, E.; and Lample, G. 2023 a . LLaMA: Open and Efficient Foundation Language Models. arXiv:2302.13971

  25. [33]

    C.; Chen, M.; Cucurull, G.; Esiobu, D.; Fernandes, J.; Fu, J.; et al

    Touvron, H.; Martin, L.; Stone, K.; Albert, P.; Almahairi, A.; Babaei, Y.; Bashlykov, N.; Batra, S.; Bhargava, P.; Bhosale, S.; Bikel, D.; Blecher, L.; Ferrer, C. C.; Chen, M.; Cucurull, G.; Esiobu, D.; Fernandes, J.; Fu, J.; et al. 2023 b . Llama 2: Open Foundation and Fine-T...

  26. [34]

    Wang, K.; Pan, J.; Shi, W.; Lu, Z.; Ren, H.; Zhou, A.; Zhan, M.; and Li, H. 2024. Measuring multimodal mathematical reasoning with math-vision dataset. Advances in Neural Information Processing Systems, 37: 95095--95169

  27. [35]

    Wang, P.; Li, L.; Shao, Z.; Xu, R.; Dai, D.; Li, Y.; Chen, D.; Wu, Y.; and Sui, Z. 2023. Math-shepherd: Verify and reinforce llms step-by-step without human annotations. CoRR, abs/2312.08935

  28. [36]

    Wang, S.; Liu, Z.; Wei, J.; Yin, X.; Li, D.; and Barsoum, E. 2025 a . Athena: Enhancing Multimodal Reasoning with Data-efficient Process Reward Models. arXiv:2506.09532

  29. [37]

    Wang, W.; Gao, Z.; Chen, L.; Chen, Z.; Zhu, J.; Zhao, X.; Liu, Y.; Cao, Y.; Ye, S.; Zhu, X.; et al. 2025 b . Visualprm: An effective process reward model for multimodal reasoning. arXiv preprint arXiv:2503.10291

  30. [38]

    Yan, Y.; and Lee, J. 2024. Georeasoner: Reasoning on geospatially grounded context for natural language understanding. In Proceedings of the 33rd ACM international conference on information and knowledge management, 4163--4167

  31. [39]

    Yan, Y.; Su, J.; He, J.; Fu, F.; Zheng, X.; Lyu, Y.; Wang, K.; Wang, S.; Wen, Q.; and Hu, X. 2024 a . A survey of mathematical reasoning in the era of multimodal large language model: Benchmark, method & challenges. arXiv preprint arXiv:2412.11936

  32. [40]

    Yan, Y.; Wang, S.; Huo, J.; Li, H.; Li, B.; Su, J.; Gao, X.; Zhang, Y.-F.; Xu, T.; Chu, Z.; et al. 2024 b . Errorradar: Benchmarking complex mathematical reasoning of multimodal large language models via error detection. arXiv preprint arXiv:2410.04509

  33. [41]

    S.; Gomes, C.; Selman, B.; and Wen, Q

    Yan, Y.; Wang, S.; Huo, J.; Ye, J.; Chu, Z.; Hu, X.; Yu, P. S.; Gomes, C.; Selman, B.; and Wen, Q. 2025 a . Position: Multimodal large language models can significantly advance scientific reasoning. arXiv preprint arXiv:2502.02871

  34. [42]

    S.; Hu, X.; and Wen, Q

    Yan, Y.; Wang, S.; Huo, J.; Yu, P. S.; Hu, X.; and Wen, Q. 2025 b . Mathagent: Leveraging a mixture-of-math-agent framework for real-world multimodal mathematical error detection. arXiv preprint arXiv:2503.18132

  35. [43]

    Yan, Y.; Wen, H.; Zhong, S.; Chen, W.; Chen, H.; Wen, Q.; Zimmermann, R.; and Liang, Y. 2024 c . Urbanclip: Learning text-enhanced urban region profiling with contrastive language-image pretraining from the web. In Proceedings of the ACM Web Conference 2024, 4006--4017

  36. [44]

    Yao, Y.; Yu, T.; Zhang, A.; Wang, C.; Cui, J.; Zhu, H.; Cai, T.; Li, H.; Zhao, W.; He, Z.; Chen, Q.; Zhou, H.; Zou, Z.; Zhang, H.; Hu, S.; Zheng, Z.; Zhou, J.; Cai, J.; Han, X.; Zeng, G.; Li, D.; Liu, Z.; and Sun, M. 2024. MiniCPM-V: A GPT-4V Level MLLM on Your Phone. arXiv:2408.01800

  37. [45]

    Yuan, L.; Li, W.; Chen, H.; Cui, G.; Ding, N.; Zhang, K.; Zhou, B.; Liu, Z.; and Peng, H. 2024. Free process rewards without process labels. arXiv preprint arXiv:2412.01981

  38. [46]

    Zeng, T.; Zhang, S.; Wu, S.; Classen, C.; Chae, D.; Ewer, E.; Lee, M.; Kim, H.; Kang, W.; Kunde, J.; et al. 2025. Versaprm: Multi-domain process reward model via synthetic reasoning data. arXiv preprint arXiv:2502.06737

  39. [47]

    Zhang, K.; Zhang, J.; Li, H.; Zhu, X.; Hua, E.; Lv, X.; Ding, N.; Qi, B.; and Zhou, B. 2025 a . Openprm: Building open-domain process-based reward models with preference trees. In The Thirteenth International Conference on Learning Representations

  40. [48]

    Zhang, R.; Jiang, D.; Zhang, Y.; Lin, H.; Guo, Z.; Qiu, P.; Zhou, A.; Lu, P.; Chang, K.-W.; Qiao, Y.; et al. 2024. Mathverse: Does your multi-modal llm truly see the diagrams in visual math problems? In European Conference on Computer Vision, 169--186. Springer

  41. [49]

    Zhang, Z.; Zheng, C.; Wu, Y.; Zhang, B.; Lin, R.; Yu, B.; Liu, D.; Zhou, J.; and Lin, J. 2025 b . The Lessons of Developing Process Reward Models in Mathematical Reasoning. arXiv:2501.07301

  42. [50]

    Zhao, J.; Liu, R.; Zhang, K.; Zhou, Z.; Gao, J.; Li, D.; Lyu, J.; Qian, Z.; Qi, B.; Li, X.; and Zhou, B. 2025. GenPRM: Scaling Test-Time Compute of Process Reward Models via Generative Reasoning. arXiv:2504.00891

  43. [51]

    Zheng, C.; Zhang, Z.; Zhang, B.; Lin, R.; Lu, K.; Yu, B.; Liu, D.; Zhou, J.; and Lin, J. 2024 a . Processbench: Identifying process errors in mathematical reasoning. arXiv preprint arXiv:2412.06559

  44. [52]

    Zheng, K.; Chen, J.; Yan, Y.; Zou, X.; and Hu, X. 2024 b . Reefknot: A comprehensive benchmark for relation hallucination evaluation, analysis and mitigation in multimodal large language models. arXiv preprint arXiv:2408.09429

  45. [53]

    Zhong, J.; Shen, W.; Li, Y.; Gao, S.; Lu, H.; Chen, Y.; Zhang, Y.; Zhou, W.; Gu, J.; and Zou, L. 2025. A comprehensive survey of reward models: Taxonomy, applications, challenges, and future. arXiv preprint arXiv:2504.12328

  46. [54]

    Y.; Dodge, J.; Fang, A.; Yu, Y.; Schmidt, L.; Wang, W

    Zhu, W.; Hessel, J.; Awadalla, A.; Gadre, S. Y.; Dodge, J.; Fang, A.; Yu, Y.; Schmidt, L.; Wang, W. Y.; and Choi, Y. 2023. Multimodal C4: An Open, Billion-scale Corpus of Images Interleaved with Text. arXiv:2304.06939

  47. [55]

    Zhuang, W.; Huang, X.; Zhang, X.; and Zeng, J. 2025. Math-puma: Progressive upward multimodal alignment to enhance mathematical reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, 26183--26191

  48. [56]

    Zou, C.; Guo, X.; Yang, R.; Zhang, J.; Hu, B.; and Zhang, H. 2024. Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models. arXiv preprint arXiv:2411.00836

Pith tools

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