Pith. sign in

REVIEW 6 major objections 8 minor 34 references

EduFlow: Advancing MLLMs' Problem-Solving Proficiency through Multi-Stage, Multi-Perspective Critique

T0 review · 6 major / 8 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 A process-reward model, reused across all pipeline stages, lifts K-12 STEM reasoning by 8.2%.

desk verdict A solid empirical pipeline for improving MLLM reasoning in K-12 STEM, but the process-oriented gains may be partly self-referential because the same reward model labels, filters, and scores the steps. read the letter →

arxiv 2507.09374 v1 pith:45VGBZPG submitted 2025-07-12 cs.AI

Chenglin Zhu , Tao Zhang , Chong Li , Mingan Lin , Zenan Zhou , Jian Xie This is my paper
classification cs.AI
keywords processrewardmodelmultimodallargelanguagemodelsMonteCarloTreeSearcheducationalreasoningK-12STEMbenchmarksbest-of-Ninferenceself-reflectiondatacuration
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

EduFlow sets out to prove that the weak spot in multimodal large language models on K-12 science is not just intelligence but supervision: models lack step-level, process-aware feedback, so errors compound across long reasoning chains. The paper claims that one process reward model, EduPRM, can supply that feedback at every stage—filtering weak training data, guiding Monte Carlo Tree Search, and reranking sampled answers—and that as a complete pipeline EduFlow raises both final-answer accuracy and step-by-step reasoning quality. Across four Qwen model sizes and four benchmarks, the reported gains are consistent, with process-oriented and result-oriented metrics both up 8.2% on the main benchmark. A reader should care because the pipeline is model-agnostic in principle: if true, it offers a reusable recipe for turning noisy, unstructured reasoning traces into structured supervision for any scientific reasoning domain.

What carries the argument

The central mechanism is EduPRM, a process-aware reward model that assigns each reasoning step a score, one of nine error labels, and a written justification, paired with EduMCTS, a Monte Carlo Tree Search whose nodes are six pedagogical actions: caption, summary, sub_task, thinking, self-reflection, and answer. EduPRM's feedback prunes and weights search branches, filters the resulting EduMCTS-160K trajectories, and reranks sampled answers during best-of-N inference; that single reward model is the mechanism carrying all three stages.

What would settle it

Use the K12Vista test set and replace EduPRM's accumulated step scores with random permutations of the same scores in best-of-N selection; if step-by-step accuracy stays at 43.28, the reranker is not doing causal work and the central claim collapses.

Watch

Extended reading notes

Core claim

The paper presents EduFlow as an end-to-end framework whose central claim is that a single process-aware reward model can serve as selector, critic, and verifier throughout the whole reasoning pipeline. EduPRM critiques each reasoning step with a score, one of nine pedagogical error labels, and a justification, and it is trained via curriculum learning on three complementary supervision sources: MCTS-guided trajectories, error-injected critiques, and teacher-student dialogues. EduMCTS then uses EduPRM's fine-grained feedback to guide a Monte Carlo Tree Search with six pedagogical action types, including self-reflection, producing the filtered EduMCTS-160K training dataset. On the paper's own terms, plugging EduPRM and EduMCTS together achieves state-of-the-art performance on K12Vista and consistent gains on MDK12, MathVision, and MMMU-Pro-V-COT, with process-oriented and result-oriented metrics both improving by 8.2%.

Load-bearing premise

The central assumption is that EduPRM's step labels are accurate enough that every stage—filtering, search, and reranking—selects genuinely better reasoning rather than outputs that merely fit EduPRM's learned format.

Editorial extensions

If this is right

  • Fine-tuning on EduMCTS-160K raises both direct and step-by-step accuracy on K12Vista across all four tested Qwen model sizes.
  • Best-of-N inference reranked by EduPRM keeps improving from N=2 to N=12, whereas random selection, self-consistency, and general-purpose ORM/PRM verifiers saturate or degrade.
  • The trained models transfer to MDK12, MathVision, and MMMU-Pro-V-COT, with Qwen2.5-VL-7B rising from 26.87 to 30.46 average accuracy.
  • EduMCTS's success rate improves from 67.8% with vanilla MCTS to 87.1% with stepwise action nodes and the EduPRM judge.
  • On K12Vista, EduFlow's best variant with EduMCTS-160K plus best-of-N (N=8) surpasses most open-source and closed-source comparison models in both direct and step-by-step reasoning.

Reading between the lines

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

  • The paper's 7B EduPRM outperforms a 72B general-purpose verifier in best-of-N selection, suggesting that a compact task-specialized process model can substitute for much larger rerankers; this is an inference, since the paper does not test other small specialized PRMs.
  • Because data selection, search, and reranking all consume the same label taxonomy, the reported process gains could partly reflect agreement with EduPRM's format rather than human-valued correctness; the paper's own filtering guarantee would need an independent human-labeled audit.
  • The reward-variance data selection step is a general recipe for cleaning noisy chain-of-thought corpora, so it could be applied to other multimodal reasoning domains, such as diagram-heavy engineering or medical exams, even without educational curricula.
  • A testable extension is to train EduPRM on fully independent human step labels and check whether the +8.2% process gain survives; if it does, the loop between EduMCTS trajectories and EduPRM training is not the source of the improvement.
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

6 major / 8 minor

Summary. EduFlow proposes a full-stack pipeline for improving multimodal large language models (MLLMs) on K-12 scientific reasoning, centered on EduPRM, a process-aware reward model that labels reasoning steps with one of nine error categories and provides justifications. EduPRM is trained on EduPRM-420K, which includes MCTS-guided trajectories, error-injected critiques, and teacher-student dialogues, and is then used in three roles: filtering raw trajectories, guiding a domain-adapted Monte Carlo Tree Search (EduMCTS) that produces the EduMCTS-160K training set, and performing best-of-N reranking at inference. The authors fine-tune four Qwen-family MLLMs on EduMCTS-160K and report consistent gains on K12Vista (Table 1) and three other benchmarks (Table 2), with additional ablations on the contribution of each stage. The central claim is that integrating EduMCTS and EduPRM yields state-of-the-art performance and an 8.2% improvement in both process-oriented and result-oriented metrics.

Significance. If the empirical claims are correct, EduFlow would provide a reusable, data-centric recipe for improving step-level reasoning in MLLMs on educational STEM tasks, with a sizeable public dataset (EduMCTS-160K) and a unified process-reward backbone spanning data selection, search, and inference. The paper's strengths are the breadth of the evaluation (four model sizes, four benchmarks, ablations of search components) and the apparently consistent direct-answer improvements. However, the process-oriented evaluation is potentially self-referential because EduPRM is used both to construct the training data and to score the step-level metrics, and the paper does not report uncertainty quantification or contamination checks. The limitations section honestly notes the narrow range of base models and domains, but it does not address the circularity in the evaluation loop.

major comments (6)
  1. [Sections 3.1.1, 3.3, and 4.2] The construction of EduPRM-420K includes 150K MCTS-guided trajectories generated by EduMCTS, yet EduMCTS (Section 3.3) uses EduPRM as its reward model and critic. The paper does not explain how the initial EduPRM was obtained to run EduMCTS for generating those 150K trajectories, resulting in an unresolved bootstrap circularity. Moreover, Section 3.3 filters EduMCTS-160K to trajectories in which all steps are labeled correct by EduPRM, and Section 3.2.3 uses EduPRM for best-of-N reranking at inference. Section 4.2 does not state which judge computes the Step-by-Step Score in Table 1. If EduPRM or a model fine-tuned on the same nine-label taxonomy produces those scores, the process-oriented gains (e.g., 27.17 to 43.28 for Qwen2.5-VL-7B) are not independent evidence of better reasoning. Please clarify the training order and the evaluation judge, and report step-level scores from an independent judge (e.g., human annotation or a held-out model not trained on this taxonomy).
  2. [Section 4.2 and Tables 1-2] No error bars, confidence intervals, or significance tests are reported for any of the main results. The average gains in Table 2 are small (e.g., +3.59 average for Qwen2.5-VL-7B), so the claim of "consistent improvements" across model sizes and benchmarks is not statistically supported. Please report standard deviations across at least three independent runs (or seeds), or a paired significance test, and state the number of test instances per benchmark.
  3. [Sections 3.1.1 and 4.2] The raw corpus for EduPRM-420K consists of 160K multimodal science problems from scanned exams, online platforms, and structured question banks. K12Vista is a benchmark introduced by the same research group (Li et al., 2025), and the paper does not report any contamination check between the training corpora (EduPRM-420K / EduMCTS-160K) and the evaluation benchmarks (K12Vista, MDK12, MathVision, MMMU-Pro-V-COT). Please report overlap statistics (e.g., n-gram overlap or embedding similarity) and per-benchmark contamination analysis.
  4. [Abstract and Introduction] The claims of "state-of-the-art performance" and "consistent improvements in both process-oriented (+8.2%) and result-oriented (+8.2%) metrics" are not supported by the reported tables. Table 1 shows Gemini2-Thinking achieving an average score of 56.42, substantially above EduFlow's best average of 44.61. Recomputing average gains from Tables 1 and 2 yields values different from 8.2% (e.g., average direct-accuracy gain of 6.87 points and average step-score gain of 16.09 points on K12Vista). Please define how the +8.2% figure is computed, and revise the SOTA claim to accurately reflect the comparisons (e.g., "competitive with or better than open-source models of similar scale").
  5. [Introduction] The sentence "EduMCTS-160K achives an 18% higher success rate than LLaVA-CoT in science domains" is unsupported: no comparison to LLaVA-CoT appears anywhere in the paper, and Table 6 reports success rates for vanilla MCTS (67.8%) versus the best EduMCTS component configuration (87.1%), not for LLaVA-CoT. Please remove this claim or substantiate it with a direct comparison to LLaVA-CoT under the same evaluation setup.
  6. [Section 3.1.1 and Appendix A.5] The step-label taxonomy is internally inconsistent. The text defines nine labels, including "Off-topic or Incongruent" (Section 3.1.1 and A.3), but Figure 2 says "covering 8 error types," the first prompt in Figure 5 lists only eight labels while stating "9 elements," and the second prompt in Figure 6 lists nine labels. Since the taxonomy is the core supervision signal for EduPRM and the apparent basis for the process evaluation, these inconsistencies must be resolved and a single canonical label set presented.
minor comments (8)
  1. [Section 1] The phrase "by unifying all stages under, EduFlow provides" is an incomplete sentence; it should read "under EduPRM" or similar.
  2. [Section 7.1] The sentence "The base model was tested only on Qwen2VL and Qwen25VL, without testing on additional base models" is duplicated verbatim.
  3. [References] Several references are duplicated: Guo et al. 2025a and 2025b are the same paper, Yang et al. 2024a and 2024b are the same, and Xu et al. 2024 and 2025 are the same LLaVA-CoT paper. Please consolidate.
  4. [Table 3] The model name "Qwen2-VL-72b" is inconsistently capitalized; elsewhere it is "Qwen2-VL-72B".
  5. [Section 3.1.1 and 3.1.2] Section 3.1.1 describes each step annotation as a triple {content, label, explanation}, while Section 3.1.2 says the quadruple is {content, label, explanation, score}. Please clarify when the score is added and what values it takes.
  6. [Appendix A.5] In the first prompt (Figure 5), the instruction says to choose from a "9元素列表" (nine-element list) but only eight labels are listed; the same prompt labels the annotations as an 8-element list in the preceding sentence. Please correct the count and ensure consistency with the canonical label set.
  7. [General] The terms "EduMCTS16W" and "EduMCTS-160K" are used interchangeably in Table 1 and elsewhere; please define the abbreviation and use it consistently.
  8. [General] There are several typos, including "achives" in the Introduction, "subejects" in Appendix A.1, and "MathVison" in Table 2. A full proofread is recommended.

Circularity Check

2 steps flagged · score 5.0 of 10

Partial circularity: EduPRM trains on EduMCTS rollouts while EduMCTS is guided and filtered by EduPRM; reported success-rate and process metrics are scored with EduPRM's own labels, though direct-answer gains on external benchmarks remain independent.

  1. self definitional [Section 3.1.1 (EduPRM-420K), Section 3.2.2 (PRM-Guided MCTS Rollout), Section 3.3 (EduMCTS)]
    "we construct EduPRM-420K using three complementary strategies ... (1) MCTS-guided trajectories (150K), where selected problems are processed by EduMCTS to generate verified, multi-step solutions through structured search; ... During trajectory construction with MCTS, EduPRM is used as the reward function to evaluate intermediate states. ... After each rollout, EduPRM evaluates the entire trajectory and assigns a reward based on step-level correctness, coherence, and the presence of identifiable errors."

    EduPRM's training set includes 150K trajectories produced by EduMCTS, while EduMCTS's rollouts, reward signals, and dataset curation are driven by EduPRM's step-level scores and 'correct' labels. The paper does not specify an independent seed judge for the MCTS-guided 150K subset, so as written the step-level ground truth of EduPRM and the 'verified' status of EduMCTS-160K are mutually defined: EduPRM is trained on EduMCTS output, and EduMCTS quality is judged by EduPRM. Consequently, claims that EduPRM is a reliable process judge and that EduMCTS-160K contains 'verified' reasoning are supported by the same loop rather than by an external anchor.

  2. fitted input called prediction [Section 3.3 (EduMCTS filtering) and Section 5.2 (Table 6)]
    "To ensure output quality, we apply a two-stage filtering process. First, only trajectories in which all steps are labeled as correct are retained. ... Table 6 shows that success rate improves from 67.8% (vanilla MCTS) to 73.3% (+actions) and 87.1% (+EduPRM), indicating better convergence with fewer invalid or redundant paths."

    The reported success rate is the proportion of rollouts that survive the 'all steps labeled as correct' filter, where the labels are produced by EduPRM. Presenting the rise to 87.1% as evidence that EduPRM supervision yields 'better convergence with fewer invalid or redundant paths' is partly circular: the notion of 'invalid' is defined by EduPRM's own step labels. The improvement therefore reflects, at least in part, EduPRM approving outputs that match its learned label format, not an independent measure of trajectory quality.

full rationale

The paper's result-oriented claims are substantially supported by independent final-answer benchmarks: Table 2 reports gains on MDK12, MathVision, and MMMU-Pro-V-COT, where correctness is checked against ground-truth answers, and Table 4's BoN improvements are measured by direct accuracy on K12Vista. These results do not reduce to EduPRM's own labels by construction, so the paper is not wholly circular. However, the process-oriented side of the paper contains a load-bearing self-referential loop. EduPRM-420K includes 150K MCTS-guided trajectories from EduMCTS, while EduMCTS uses EduPRM as its reward model and judge, and EduMCTS-160K is filtered by retaining only trajectories 'in which all steps are labeled as correct' by EduPRM. The paper does not state an independent seed or external ground-truth anchor for the step labels in this loop. In addition, the success-rate ablation in Table 6 is defined by EduPRM's own approvals, making the reported improvement from 67.8% to 87.1% partly a measure of self-consistency rather than an external quality metric. The Step-by-Step evaluation in Table 1 and Figure 1 is not specified as coming from an independent judge; if it uses EduPRM or a model trained on EduPRM's nine-label taxonomy, the process-oriented '+8.2%' claim would largely be a self-referential score. Weighing these issues, the mutual bootstrap and label-self-consistency create partial circularity, but the independent direct-answer results prevent a score of 8 or 10. A score of 5 reflects that the core result-oriented claims have independent content while the process-oriented evidence is not fully independent.

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

EduPRM and EduMCTS are trained artifacts with direct benchmark evaluation, not unobservable postulated entities; no new particles, forces, or non-falsifiable constructs are introduced.

free parameters (4)
  • MCTS pruning threshold tau = not disclosed
    Discards low-scoring steps during EduMCTS expansion (Appendix A.2); the quality of filtered trajectories depends on this value.
  • UCB exploration coefficient c = not disclosed
    Balances exploration and exploitation in node selection (Appendix A.2).
  • MCTS rollout count = 4
    Ablation in Table 5 shows step score improves from 25.64 to 39.50 as rollouts increase from 1 to 4, then saturates at 8.
  • Best-of-N sample size = 8
    Table 4 shows accuracy rises from 41.21 (N=2) to 43.28 (N=8) and barely improves at N=12; N=8 is the chosen operating point.
assumptions (4)
  • domain assumption The nine step-level error labels are a complete and valid taxonomy for educational reasoning quality.
    Used throughout for data annotation (EduPRM-420K) and step-by-step evaluation (Appendix A.3).
  • domain assumption LLM-generated annotations (GPT-4o error injection, Qwen2.5-VL-72B teacher labels, and EduPRM self-labels) are treated as ground truth after format and consistency filtering.
    Section 3.1.1 and Appendix A.4 describe automatic filtering, not human verification of correctness.
  • domain assumption The 160K training problems do not overlap with the evaluation benchmarks.
    No contamination analysis is reported despite similar sourcing from scanned exams and online platforms (Section 3.1.1).
  • domain assumption The step-by-step evaluation score reflects genuine reasoning quality rather than agreement with EduPRM's step format.
    The process metric uses the same nine-label taxonomy on which EduPRM was trained (Section 5.1), so format alignment could inflate the score.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EduFlow: Advancing MLLMs' Problem-Solving Proficiency through Multi-Stage, Multi-Perspective Critique." pith.science (2026). https://pith.science/paper/45VGBZPG

@misc{pith2026250709374,
  author       = {Pith},
  title        = {Pith review of: EduFlow: Advancing MLLMs' Problem-Solving Proficiency through Multi-Stage, Multi-Perspective Critique},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/45VGBZPG}},
  note         = {Machine review of arXiv:2507.09374}
}
read the original abstract

Multimodal large language models (MLLMs) still perform poorly on scientific tasks, particularly those requiring multi-step and interpretable reasoning. Their limitations include insufficient scientific reasoning patterns, lack of global coherence in multi-step inference, and the absence of reflective self-correction, making them unreliable in structured scientific contexts. We introduce EduFlow, the first end-to-end framework that covers the full pipeline of educational scientific reasoning, including data selection, MCTS-based trajectory construction, model training, and output optimization. At its core is EduPRM, a process-aware reward model that critiques reasoning steps with tags and justifications. EduPRM is trained via curriculum learning on three complementary supervision sources: MCTS-guided trajectories, error-injected critiques, and teacher-student dialogues, enabling dynamic adaptation to multi-stage problem solving and iterative refinement during inference. We further propose EduMCTS, a domain-adapted search framework that introduces bootstrapping actions specifically designed for educational reasoning, such as a self-reflection mechanism that promotes reflective error correction. It further leverages EduPRM's fine-grained feedback to guide the search toward higher-quality reasoning trajectories. By applying self-consistency and rejection sampling, we constructed EduMCTS-160K, a large-scale dataset of educational reasoning trajectories. Extensive experiments demonstrate that EduFlow enhances reasoning consistency and coherence. Code, data, and models will be released.

Figures

Figures reproduced from arXiv: 2507.09374 by the authors.

Figure 1
Figure 1. Step-by-step reasoning performance across [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the construction and annotation pipeline for [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Overview of the multi-stage usage of EduPRM in the EduFlow. EduPRM plays distinct roles at each stage: (1) as a Selector to filter CoT outputs during Raw CoT Generation, enabling data selection; (2) as a critic to evaluate and provide stepwise feedback on rollouts during EduMCTS Exploration, guiding reward propagation and effectively search; (3) as a verifier to select the best answer among sampled responses during … view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: presents subject-wise accuracy compar￾isons across core K-12 domains. Our method con￾sistently improves performance across all subjects, with particularly significant gains in STEM fields such as Math, Physics, and Chemistry. MCQ FBQ FRQ Math Physics Chemistry Biology …
Figure 5
Figure 5. Figure 5: Prompt for step breakdown and annotation [PITH_FULL_IMAGE:figures/full_fig_p013_5.png]
Figure 6
Figure 6. Figure 6: Prompt for step annotation [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

34 extracted references · 7 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, and 1 others. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  4. [4]

    Anthropic. 2024. Claude 3.5 sonnet

    work page 2024

  5. [5]

    Byrd, Robert Zinkov, and Nada Amin

    David Brandfonbrener, Simon Henniger, Sibi Raja, Tarun Prasad, Chloe Loughridge, Federico Cassano, Sabrina Ruixin Hu, Jianang Yang, William E. Byrd, Robert Zinkov, and Nada Amin. 2024. https://arxiv.org/abs/2402.08147 Vermcts: Synthesizing multi-step programs using a verifier, a large language model, and tree search . Preprint, arXiv:2402.08147

    arXiv 2024

  6. [6]

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, and 1 others. 2024. Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling. arXiv preprint arXiv:2412.05271

    arXiv 2024

  7. [7]

    OpenCompass Contributors. 2023. Opencompass: A universal evaluation platform for foundation models. https://github.com/open-compass/opencompass

    2023

  8. [8]

    Guanting Dong, Chenghao Zhang, Mengjie Deng, Yutao Zhu, Zhicheng Dou, and Ji-Rong Wen. 2024. https://arxiv.org/abs/2412.14835 Progressive multimodal reasoning via active retrieval . Preprint, arXiv:2412.14835

    arXiv 2024

Show all 34 references
  1. [9]

    Zitian Gao, Boye Niu, Xuzheng He, Haotian Xu, Hongzhang Liu, Aiwei Liu, Xuming Hu, and Lijie Wen. 2024. https://arxiv.org/abs/2410.01707 Interpretable contrastive monte carlo tree search reasoning . Preprint, arXiv:2410.01707

    2024 arXiv

  2. [10]

    Xinyu Guan, Li Lyna Zhang, Yifei Liu, Ning Shang, Youran Sun, Yi Zhu, Fan Yang, and Mao Yang. 2025. https://arxiv.org/abs/2501.04519 rstar-math: Small llms can master math reasoning with self-evolved deep thinking . Preprint, arXiv:2501.04519

    2025 arXiv

  3. [12]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, and 1 others. 2025 b . Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    2025 arXiv

  4. [13]

    Zheqi He, Xinya Wu, Pengfei Zhou, Richeng Xuan, Guang Liu, Xi Yang, Qiannan Zhu, and Hua Huang. 2024. https://arxiv.org/abs/2401.14011 Cmmu: A benchmark for chinese multi-modal multi-type question understanding and reasoning . Preprint, arXiv:2401.14011

    2024 arXiv

  5. [14]

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, and 1 others. 2024. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326

    2024 arXiv

  6. [15]

    Chong Li, Chenglin Zhu, Tao Zhang, Mingan Lin, Zenan Zhou, and Jian Xie. 2025. https://arxiv.org/abs/2506.01676 K12vista: Exploring the boundaries of mllms in k-12 education . Preprint, arXiv:2506.01676

    2025 arXiv

  7. [16]

    Hunter Lightman, Vineet Kosaraju, Yura Burda, Harri Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. 2023. https://arxiv.org/abs/2305.20050 Let's verify step by step . Preprint, arXiv:2305.20050

    2023 arXiv

  8. [17]

    Shuhang Liu, Zhenrong Zhang, Pengfei Hu, Jiefeng Ma, Jun Du, Qing Wang, Jianshu Zhang, Quan Liu, Jianqing Gao, and Feng Ma. 2025. https://arxiv.org/abs/2504.11009 Mmc: Iterative refinement of vlm reasoning via mcts-based multimodal critique . Preprint, arXiv:2504.11009

    2025 arXiv

  9. [18]

    OpenAI. 2024 a . Gpt-4v(ision) system card

    2024

  10. [19]

    OpenAI. 2024 b . Gpt-o1-mini: A multimodal language model

    2024

  11. [20]

    OpenAI. 2024 c . Gpt-o3-mini: A cost-effective reasoning model

    2024

  12. [21]

    qwen Team. 2024. Qvq: To see the world with wisdom

    2024

  13. [22]

    Z. Z. Ren, Zhihong Shao, Junxiao Song, Huajian Xin, Haocheng Wang, Wanjia Zhao, Liyue Zhang, Zhe Fu, Qihao Zhu, Dejian Yang, Z. F. Wu, Zhibin Gou, Shirong Ma, Hongxuan Tang, Yuxuan Liu, Wenjun Gao, Daya Guo, and Chong Ruan. 2025. https://arxiv.org/abs/2504.21801 Deepseek-prove...

    2025 arXiv

  14. [23]

    Jianhao Shen, Ye Yuan, Srbuhi Mirzoyan, Ming Zhang, and Chenguang Wang. 2024. https://arxiv.org/abs/2402.17205 Measuring vision-language stem skills of neural models . Preprint, arXiv:2402.17205

    2024 arXiv

  15. [24]

    Jonathan Uesato, Nate Kushman, Ramana Kumar, Francis Song, Noah Siegel, Lisa Wang, Antonia Creswell, Geoffrey Irving, and Irina Higgins. 2022. https://arxiv.org/abs/2211.14275 Solving math word problems with process- and outcome-based feedback . Preprint, arXiv:2211.14275

    2022 arXiv

  16. [25]

    Hongyu Wang, Jiayu Xu, Senwei Xie, Ruiping Wang, Jialin Li, Zhaojie Xie, Bin Zhang, Chuyan Xiong, and Xilin Chen. 2024 a . https://arxiv.org/abs/2405.15638 M4u: Evaluating multilingual understanding and reasoning for large multimodal models . Preprint, arXiv:2405.15638

    2024 arXiv

  17. [26]

    Weiyun Wang, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, Xizhou Zhu, Lewei Lu, Yu Qiao, and 1 others. 2024 b . Enhancing the reasoning ability of multimodal large language models via mixed preference optimization. arXiv preprint arXiv:2411.10442

    2024 arXiv

  18. [27]

    Weiyun Wang, Zhangwei Gao, Lianjie Chen, Zhe Chen, Jinguo Zhu, Xiangyu Zhao, Yangzhou Liu, Yue Cao, Shenglong Ye, Xizhou Zhu, Lewei Lu, Haodong Duan, Yu Qiao, Jifeng Dai, and Wenhai Wang. 2025. https://arxiv.org/abs/2503.10291 Visualprm: An effective process reward model for m...

    2025 arXiv

  19. [29]

    Guowei Xu, Peng Jin, Hao Li, Yibing Song, Lichao Sun, and Li Yuan. 2025. https://arxiv.org/abs/2411.10440 Llava-cot: Let vision language models reason step-by-step . Preprint, arXiv:2411.10440

    2025 arXiv

  20. [31]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, and 1 others. 2024 b . Qwen2. 5 technical report. arXiv preprint arXiv:2412.15115

    2024 arXiv

  21. [32]

    Huanjin Yao, Jiaxing Huang, Wenhao Wu, Jingyi Zhang, Yibo Wang, Shunyu Liu, Yingjie Wang, Yuxin Song, Haocheng Feng, Li Shen, and Dacheng Tao. 2024. https://arxiv.org/abs/2412.18319 Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree...

    2024 arXiv

  22. [33]

    Dan Zhang, Sining Zhoubian, Ziniu Hu, Yisong Yue, Yuxiao Dong, and Jie Tang. 2024 a . https://arxiv.org/abs/2406.03816 Rest-mcts*: Llm self-training via process reward guided tree search . Preprint, arXiv:2406.03816

    2024 arXiv

  23. [34]

    Di Zhang, Xiaoshui Huang, Dongzhan Zhou, Yuqiang Li, and Wanli Ouyang. 2024 b . https://arxiv.org/abs/2406.07394 Accessing gpt-4 level mathematical olympiad solutions via monte carlo tree self-refine with llama-3 8b . Preprint, arXiv:2406.07394

    2024 arXiv

  24. [35]

    Ge Zhang, Xinrun Du, Bei Chen, Yiming Liang, Tongxu Luo, Tianyu Zheng, Kang Zhu, Yuyang Cheng, Chunpu Xu, Shuyue Guo, Haoran Zhang, Xingwei Qu, Junjie Wang, Ruibin Yuan, Yizhi Li, Zekun Wang, Yudong Liu, Yu-Hsuan Tsai, Fengji Zhang, and 3 others. 2024 c . https://arxiv.org/abs...

    2024 arXiv

  25. [36]

    Jian Zhao, Runze Liu, Kaiyan Zhang, Zhimu Zhou, Junqi Gao, Dong Li, Jiafei Lyu, Zhouyi Qian, Biqing Qi, Xiu Li, and Bowen Zhou. 2025. https://arxiv.org/abs/2504.00891 Genprm: Scaling test-time compute of process reward models via generative reasoning . Preprint, arXiv:2504.00891

    2025 arXiv

  26. [37]

    Pengfei Zhou, Fanrui Zhang, Xiaopeng Peng, Zhaopan Xu, Jiaxin Ai, Yansheng Qiu, Chuanhao Li, Zhen Li, Ming Li, Yukang Feng, Jianwen Sun, Haoquan Zhang, Zizhen Li, Xiaofeng Mao, Wangbo Zhao, Kai Wang, Xiaojun Chang, Wenqi Shao, Yang You, and Kaipeng Zhang. 2025. https://arxiv.o...

    2025 arXiv

Pith tools

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