Pith. sign in

REVIEW 4 major objections 7 minor 4 cited by

VL-Cogito: Progressive Curriculum Reinforcement Learning for Advanced Multimodal Reasoning

T0 review · 4 major / 7 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 The paper claims that a progressive easy-to-hard reinforcement curriculum with a prompt-specific length reward makes a 7B multimodal model reach best or near-best scores on most of ten benchmarks, with no cold-start supervised fine-tuning.

desk verdict A useful training recipe with a genuinely new dynamic length reward, but the headline SOTA claim rests on single-run margins that are within plausible noise. read the letter →

arxiv 2507.22607 v2 pith:A2SJWPRO submitted 2025-07-30 cs.CV cs.AIcs.CL

Ruifeng Yuan , Chenghao Xiao , Sicong Leng , Jianyu Wang , Long Li , Weiwen Xu , Hou Pong Chan , Deli Zhao
show 4 more authors
Tingyang Xu Zhongyu Wei Hao Zhang Yu Rong
This is my paper
classification cs.CVcs.AIcs.CL
keywords multimodalreasoningreinforcementlearningcurriculumGRPORLVRdynamiclengthrewarddifficultysoftweightingvision-languagemodel
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

The paper is trying to establish that curriculum structure can make reinforcement learning for multimodal reasoning both stable and efficient. It trains a 7B vision-language model directly from an instruct backbone through three RL stages—easy, medium, hard—using an online difficulty soft weighting that shapes each prompt's gradient contribution, and it reports best or near-best scores on most of ten benchmarks covering math, science, logic, charts, and general understanding. The authors argue that this works without the cold-start supervised fine-tuning used by several comparison systems. The reason to care is that if true, the recipe separates 'what to train on' from 'when,' turning difficulty and reasoning length into controllable training signals rather than fixed properties of the data.

What carries the argument

The load-bearing machinery is the pair of reward and advantage shapers used inside a three-stage group relative policy optimization (GRPO) run. First, Online Difficulty Soft Weighting multiplies each prompt's group advantage by a function $F(\mathrm{Acc})$ of the prompt's average rollout accuracy, choosing an easy, medium, or hard variant over successive stages; all three variants peak around accuracy $0.5$, the point where a question is neither trivially solved nor hopeless, so the model focuses on learnable difficulty while still weighting other prompts. Second, the Dynamic Length Reward $r_{\mathrm{len}}(y_i,y)$ is a cosine reward whose target $L_{\mathrm{avg}}$ is the average length of the correct rollouts for that prompt, so the target is per-question, changes as the policy improves, and never applies a uniform length budget; prompts with no correct rollout are pushed toward a preset maximum $L_{\mathrm{max}}$ with a downweighted advantage. The curriculum itself is three consecutive optimization phases over the same dataset: about 100 steps with easy weighting, 100 with medium weighting, and about 200 with hard weighting plus the length reward.

What would settle it

Give the length reward an external target by hand-annotating a subset of questions with the shortest correct solution length, then train the hard stage against those targets instead of $L_{\mathrm{avg}}$. If accuracy rises or stays flat while the model's self-derived targets were systematically shorter or longer than the annotated ones, the self-referential target is not finding the true reasoning budget; if the external targets hurt, the paper's mechanism is genuinely adaptive.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that guiding a model from easier to harder prompts through GRPO-based reinforcement learning is itself a powerful training signal. The resulting model, VL-Cogito, is reported to outperform its backbone on every evaluated benchmark, to exceed all reasoning-oriented baselines on six of ten, and to match or beat a strong rethinking-based system on most mathematical and scientific suites while never being forced to rethink. The dynamic length reward is the piece that lets the model compress answers to easy questions and expand them on hard ones; the curriculum is the piece that keeps that length expansion from destabilizing early training. Together, they are claimed to raise overall average accuracy from 58.1% for vanilla GRPO to 59.5%, with the biggest gains on the hardest reasoning sets.

Load-bearing premise

The load-bearing premise is that the average length of the model's own correct answers is a trustworthy sign of how much reasoning a question actually needs, because the whole length-reward mechanism treats that average as the target and would otherwise just lock in the policy's current length habits.

Editorial extensions

If this is right

  • Replication cost drops if no curated SFT chain-of-thought data is needed: a practitioner can start from an open instruct model and rely on the curriculum to elicit reasoning.
  • Prompt-level difficulty weighting can replace hard filtering of training data, so fewer samples are wasted and the same dataset can serve all curriculum stages.
  • A length reward can be made question-adaptive rather than global, which should reduce overthinking on easy multimodal questions and underthinking on hard ones.
  • The ablation claim implies curriculum and length shaping interact: using the length reward without the curriculum produces unstable training, so the ordering is part of the mechanism.

Reading between the lines

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

  • A natural stress test the paper does not run is to replace $L_{\mathrm{avg}}$ with an externally estimated 'ideal reasoning length' per question; if the model-derived target is close to that external oracle, the mechanism is learning something about the task, and if not, it is mostly learning its own bias.
  • The equations for $F$ and $r_{\mathrm{len}}$ are modality-neutral, so the same three-stage curriculum is a candidate recipe for text-only LLM reasoning; a cheap experiment would be to apply PCuRL to a math-only RL setting and compare against uniform GRPO.
  • Because the dataset is fixed across stages, the paper's curriculum is purely an emphasis schedule; one could test whether a single continuously shifted difficulty weighting reproduces the gains, which would determine whether the discrete easy, medium, and hard staging is essential.
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 / 7 minor

Summary. The paper proposes VL-Cogito, a 7B multimodal reasoning model trained from Qwen2.5-VL-Instruct-7B without a cold-start SFT phase, using a three-stage Progressive Curriculum Reinforcement Learning (PCuRL) framework. PCuRL combines an online difficulty soft weighting mechanism (ODSW), which reweights GRPO advantages according to per-prompt rollout accuracy using easy/medium/hard schedules, with a dynamic length reward (DyLR) whose target is the average length of the model's own correct responses (or a preset maximum when no response is correct). Experiments cover ten multimodal benchmarks; the authors report gains over vanilla GRPO averaging about 1.4 points, state-of-the-art or competitive results relative to several reasoning-oriented baselines, ablations of the curriculum and ODSW variants, training curves, and case studies. The central claim is that the PCuRL recipe improves multimodal reasoning without SFT and that DyLR adaptively balances reasoning depth and efficiency.

Significance. If the reported gains are robust, the paper makes a useful contribution: it demonstrates that curriculum-style difficulty weighting and prompt-adaptive length rewards can be combined in a GRPO pipeline for multimodal reasoning, using only open-source data and a standard backbone. The strengths are the broad benchmark suite, the unified re-evaluation of baselines, the component-wise ablations in Table 2, and the detailed ablation of ODSW variants in Table 3. The training-dynamics visualization in Figure 4 is a useful diagnostic. However, the main empirical claim currently rests on single-run differences of 0.5–1.4 points with no variance estimates, the selection procedure for key hyperparameters is not reported, and the DyLR target is self-referential. These issues are load-bearing for the 'state-of-the-art on 6 out of 10 benchmarks' claim and need to be addressed before the result can be considered established.

major comments (4)
  1. [Tables 1–3, Section 5.1] All accuracy numbers are single-run results with no error bars or multi-seed averages. The margins that support the paper's central claims are small: full PCuRL averages 59.5 versus 58.9 for +Curriculum and 58.6 for +DyLR (Table 2), and 58.7 for ODSW-Hard-only (Table 3); per-benchmark differences such as Geo3K 68.7 vs 67.7 and MathVista 74.8 vs 73.7 over VL-Rethinker are within plausible run-to-run variation for 7B-scale RL training. To support the claim in Section 5.2 of 'the highest performance on 6 out of 10 multimodal benchmarks', please provide multiple seeds for the main model and key ablations, or a statistical significance analysis (e.g., bootstrap over test examples).
  2. [Section 5.1, Figure 3, Figure 4] No sensitivity analysis is reported for the central hyperparameters. Stage lengths (100/100/~200 steps), reward coefficients (alpha=1, beta=0.5, gamma=1, w=0.25), the Lmax target (inferred from Figure 4 as 500 but not listed in the implementation details), and the decision to apply DyLR only in the hard stage are presented as fixed choices. The paper states that 'empirical experiments' and validation accuracy were used to set stage lengths and checkpoints, but it does not state whether the final benchmark suite was used for any of these selections or whether a separate validation split was used throughout. Because Table 3 shows that ODSW-Hard-only (58.7) and ODSW-Medium-only (58.6) are within about 1 point of the full recipe (59.5), selection effects could account for the reported advantage. Please add a sensitivity study (e.g., Lmax in {500,750,1000}, w in {0,0.25,0.5}) and an explicit statement of how validation was separated from the test benchmarks.
  3. [Eq. (5), Section 4.2.2] The dynamic length reward defines the target reasoning budget as Lavg, the average length of the current policy's correct responses. This makes the claimed 'adaptive' regulation partly tautological: if the policy's correct outputs are systematically too short or too long, the reward entrenches that bias rather than measuring task complexity. The claim that DyLR balances reasoning efficiency with correctness requires evidence that Lavg tracks an independent signal of the appropriate reasoning budget. Please ablate Lavg against static or oracle targets (e.g., lengths of correct solutions from a stronger model or human annotations), or report the correlation between Lavg and an external difficulty measure.
  4. [Table 3, Section 5.3.2] The single-stage ODSW ablations are described without stating their total number of optimization steps or whether DyLR is included. If the ODSW-only runs used fewer total steps than the three-stage pipeline (100+100+~200), the comparison in Table 3 conflates curriculum design with compute budget. Please report the total number of steps, effective sample count, and reward configuration for every row of Table 3, and ideally match compute across rows.
minor comments (7)
  1. [Table 1, Sections 1 and 5.2] Table 1 lists 11 benchmark columns including MMMU-Sci, while the text refers to 'ten different multimodal benchmarks' and '6 out of 10'; please clarify whether MMMU-Sci is counted as a separate benchmark.
  2. [Eq. (5)] The notation rlen(yi, y) uses y without defining it; please define the collection of rollouts or replace y with a clearer symbol for the prompt-level statistic.
  3. [Section 4.2.1] In the ODSW Medium definition, 'Sin(π · Acc)' should use lower-case 'sin' for consistency with the other cases.
  4. [Section 4.1] The phrase 'over 8 trails' should be 'over 8 trials'.
  5. [Table 5] In the training system prompt, 'The Assistant must carefully checkout the correctness' should be 'carefully check the correctness'.
  6. [Section 5.3.3, Figure 3] Figure 3 reports accuracy and length only graphically; please provide the underlying numerical values and, if possible, confidence intervals or per-benchmark breakdowns.
  7. [Section 5.1, Eq. (5)] The preset maximum length Lmax is used in Eq. (5) but never listed in the implementation details; Figure 4 indicates a value of 500 tokens, and this should be stated explicitly.

Circularity Check

1 steps flagged · score 3.0 of 10

Mild self-referential target in the dynamic length reward; central benchmark claims are empirical and not circular.

  1. self definitional [Section 4.2.2, Eq. (5)]
    "the target reasoning budget for each prompt is defined as the average length of all correct responses within its rollout samples. This strategy encourages the model to generate reasoning paths that closely align with the prompt-specific target length."

    In Eq. (5), for prompts with any correct rollouts (Acc>0), the length-reward target is Ltgt = Lavg, where Lavg is the mean length of the current policy's own correct responses for that prompt. The 'ideal' or 'prompt-specific target' length is therefore not an independent measure of task complexity; it is a statistic of the behavior being trained. Consequently, the abstract's claim that DyLR makes the model 'adaptively regulate its reasoning path length according to task complexity' is partly analytic: the reward pulls lengths toward the model's current average correct length, and the observed easy-short/hard-long pattern inherits the policy's prior length distribution rather than being independently derived from the questions.

full rationale

The paper's central claims—state-of-the-art or highly competitive accuracy on 6 of 10 external benchmarks (Table 1) and component-wise gains (Tables 2–3)—are empirical comparisons against independently trained baselines and do not reduce to any fitted parameter or self-citation. ODSW weights advantages by rollout accuracy (Eq. 4); this is a training-signal shaping rule, not a prediction derived from itself. No load-bearing uniqueness theorem or ansatz is imported from the authors' prior work; citations to MMR1 and Praxis-VLM are related-work context only. The one genuine definitional circularity is confined to the DyLR target in Eq. (5), where the per-prompt target length is defined as the average length of the model's own correct responses, making the 'adaptive length' claim partially tautological. Separately, the paper reports single runs and small ablation margins (e.g., VL-Cogito 59.5 vs ODSW-Hard-only 58.7 in Table 3) and does not report a disjoint hyperparameter-selection split; those are robustness/selection concerns, not circularity. Overall circularity is low: score 3.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method relies on several hand-chosen hyperparameters and two domain assumptions: the 0.5-accuracy optimality claim from learnability theory, and the self-referential use of correct-response length as the target budget. No new physical or conceptual entities are introduced.

free parameters (7)
  • Lmax (maximum target length) = 500 tokens in the main model; 500/750/1000 explored in Figure 3
    Preset maximum target length in Eq. (5) and Figure 4; selected by comparing Dynamic-500/750/1000 ablations, i.e., tuned on evaluation benchmarks.
  • w (zero-accuracy advantage weight) = 0.25
    Hyperparameter in Eq. (8), chosen to prevent excessive length growth for prompts with no correct rollouts.
  • Reward coefficients alpha, beta, gamma = 1, 0.5, 1
    Weights for accuracy, format, and dynamic length reward in Eq. (7), set by hand.
  • rmin_len, rmax_len = -1, 0
    Minimum and maximum length rewards in Eq. (5); length reward is a punishment.
  • Stage lengths = 100, 100, ~200 optimization steps
    Chosen based on observed reward/validation plateaus (Section 5.1), effectively fitted to validation performance.
  • Difficulty filter threshold = 50% accuracy over 8 trials
    Samples above this accuracy with Qwen2.5-VL-7B-Instruct are removed (Section 4.1); threshold chosen by hand.
  • ODSW difficulty curves = three piecewise sine/constant forms
    Hand-designed to peak at low, medium, or high accuracy (Figure 2), based on the learnability assumption.
assumptions (4)
  • domain assumption Prompts with rollout accuracy near 0.5 are optimal for RL training
    Invoked in Section 4.2.1 to justify the ODSW weighting; cited to learnability literature but not proven in this paper.
  • ad hoc to paper Average length of correct rollouts is the appropriate reasoning budget for a prompt
    Assumed in Eq. (5) of Section 4.2.2; no external benchmark or theory is given for why this statistic measures task complexity.
  • standard math GRPO with rule-based verifiable rewards applies to multimodal inputs
    Standard RLVR framework used throughout, following Shao et al.; treated as background.
  • domain assumption Removing samples solvable by a smaller instruct model removes non-reasoning questions
    Section 4.1 difficulty sampling; assumes Qwen2.5-VL-7B-Instruct accuracy is a valid difficulty oracle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of VL-Cogito: Progressive Curriculum Reinforcement Learning for Advanced Multimodal Reasoning." pith.science (2026). https://pith.science/paper/A2SJWPRO

@misc{pith2026250722607,
  author       = {Pith},
  title        = {Pith review of: VL-Cogito: Progressive Curriculum Reinforcement Learning for Advanced Multimodal Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A2SJWPRO}},
  note         = {Machine review of arXiv:2507.22607}
}
read the original abstract

Reinforcement learning has proven its effectiveness in enhancing the reasoning capabilities of large language models. Recent research efforts have progressively extended this paradigm to multimodal reasoning tasks. Due to the inherent complexity and diversity of multimodal tasks, especially in semantic content and problem formulations, existing models often exhibit unstable performance across various domains and difficulty levels. To address these limitations, we propose VL-Cogito, an advanced multimodal reasoning model trained via a novel multi-stage Progressive Curriculum Reinforcement Learning (PCuRL) framework. PCuRL systematically guides the model through tasks of gradually increasing difficulty, substantially improving its reasoning abilities across diverse multimodal contexts. The framework introduces two key innovations: (1) an online difficulty soft weighting mechanism, dynamically adjusting training difficulty across successive RL training stages; and (2) a dynamic length reward mechanism, which encourages the model to adaptively regulate its reasoning path length according to task complexity, thus balancing reasoning efficiency with correctness. Experimental evaluations demonstrate that VL-Cogito consistently matches or surpasses existing reasoning-oriented models across mainstream multimodal benchmarks spanning mathematics, science, logic, and general understanding, validating the effectiveness of our approach.

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. Full citation record

  1. Cognitive Pivot Points and Visual Anchoring: Unveiling and Rectifying Hallucinations in Multimodal Reasoning Models

    cs.AI 2026-04 unverdicted novelty 6.0 of 10

    Multimodal reasoning models hallucinate at high-entropy cognitive bifurcation points due to loss of visual semantic anchoring, and the V-STAR training paradigm with HVAR rewards and FRM reflection mitigates this by re...

  2. Wiki-R1: Incentivizing Multimodal Reasoning for Knowledge-based VQA via Data and Sampling Curriculum

    cs.CV 2026-03 conditional novelty 6.0 of 10

    Controllable retrieval-difficulty curriculum plus reward-propagation sampling lets RL close the pretrain-to-KB-VQA gap and beat prior SOTA on two hard encyclopedic VQA benchmarks.

  3. LLaVA-Critic-R1: Your Critic Model is Secretly a Strong Policy Model

    cs.CV 2025-08 conditional novelty 6.0 of 10

    RL training on preference-labeled critic data transforms a 7B vision-language model into both a stronger critic and a stronger generative policy, improving average benchmark accuracy by 5.7% and enabling self-critique...

  4. ChronoVision: Temporal Reasoning via Latent State Reconstruction

    cs.CV 2026-08 conditional novelty 5.0 of 10

    ChronoVision reaches 74.8% in-domain and 71.6% out-of-domain exact-match accuracy on the new Vbvr-VQA frame-ordering benchmark plus 55.0% on IntPhys2, with a 9B model trained on one million same-suite examples.

Reference graph

Works this paper leans on

84 extracted references · 27 canonical work pages · cited by 4 Pith papers

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    Lawrence Zitnick, and Devi Parikh

    Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C. Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), December 2015. https://openaccess.thecvf.com/content_iccv_2015/papers/Antol_VQA_Visual_Question_ICCV_2015_paper.pdf

    2015

  3. [3]

    Online difficulty filtering for reasoning oriented reinforcement learning

    Sanghwan Bae, Jiwoo Hong, Min Young Lee, Hanbyul Kim, JeongYeon Nam, and Donghyun Kwak. Online difficulty filtering for reasoning oriented reinforcement learning. ArXiv, abs/2504.03380, 2025. https://arxiv.org/abs/2504.03380

    arXiv 2025

  4. [4]

    Qwen2.5-vl technical report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report. A...

    arXiv 2025

  5. [5]

    Why reasoning matters? a survey of advancements in multimodal reasoning (v1)

    Jing Bi, Susan Liang, Xiaofei Zhou, Pinxin Liu, Junjia Guo, Yunlong Tang, Luchuan Song, Chao Huang, Guangyu Sun, Jinxi He, Jiarui Wu, Shu Yang, Daoan Zhang, Chen Chen, Lianggong Bruce Wen, Zhang Liu, Jiebo Luo, and Chenliang Xu. Why reasoning matters? a survey of advancements in multimodal reasoning (v1). ArXiv, abs/2504.03151, 2025. https://arxiv.org/abs...

    arXiv 2025

  6. [6]

    Mm-iq: Benchmarking human-like abstraction and reasoning in multimodal models

    Huanqia Cai, Yijun Yang, and Winston Hu. Mm-iq: Benchmarking human-like abstraction and reasoning in multimodal models. ArXiv, abs/2502.00698, 2025. https://arxiv.org/abs/2502.00698

    arXiv 2025

  7. [7]

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

    Jie Cao and Jing Xiao. An augmented benchmark dataset for geometric question answering through dual parallel text encoding. In Proceedings of the 29th International Conference on Computational Linguistics, pages 1511--1520, Gyeongju, Republic of Korea, October 2022. International Committee on Computational Linguistics. https://aclanthology.org/2022.coling-1.130/

    2022

  8. [8]

    Mathflow: Enhancing the perceptual flow of mllms for visual mathematical problems

    Felix Chen, Hangjie Yuan, Yunqiu Xu, Tao Feng, Jun Cen, Pengwei Liu, Zeying Huang, and Yi Yang. Mathflow: Enhancing the perceptual flow of mllms for visual mathematical problems. ArXiv, abs/2503.16549, 2025 a . https://arxiv.org/abs/2503.16549

    arXiv 2025

Show all 84 references
  1. [9]

    Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, volume 37, pages 27056--27087

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models? In Advances in Neural Information Processing Systems, volume 37, pages 27056...

    2024

  2. [10]

    Towards reasoning era: A survey of long chain-of-thought for reasoning large language models

    Qiguang Chen, Libo Qin, Jinhao Liu, Dengyun Peng, Jiannan Guan, Peng Wang, Mengkang Hu, Yuhang Zhou, Te Gao, and Wanxiang Che. Towards reasoning era: A survey of long chain-of-thought for reasoning large language models. ArXiv, abs/2503.09567, 2025 b . https://arxiv.org/abs/2503.09567

    2025 arXiv

  3. [11]

    Grpo-care: Consistency-aware reinforcement learning for multimodal reasoning

    Yi Chen, Yuying Ge, Rui Wang, Yixiao Ge, Junhao Cheng, Ying Shan, and Xihui Liu. Grpo-care: Consistency-aware reinforcement learning for multimodal reasoning. ArXiv, abs/2506.16141, 2025 c . https://arxiv.org/abs/2506.16141

    2025 arXiv

  4. [12]

    Expanding performance boundaries of open-source multimodal models with model, data, and test-time scaling

    Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, Lixin Gu, Xuehui Wang, Qingyun Li, Yimin Ren, Zixuan Chen, Jiapeng Luo, Jiahao Wang, Tan Jiang, Bo Wang, Conghui He, Botian Shi, Xingcheng Zhang, Han Lv, Yi...

    2025 arXiv

  5. [13]

    Process reinforcement through implicit rewards

    Ganqu Cui, Lifan Yuan, Zefan Wang, Hanbin Wang, Wendi Li, Bingxiang He, Yuchen Fan, Tianyu Yu, Qixin Xu, Weize Chen, et al. Process reinforcement through implicit rewards. arXiv preprint arXiv:2502.01456, 2025

    2025 arXiv

  6. [14]

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

    DeepSeek-AI. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv, abs/2501.12948, 2025. https://arxiv.org/abs/2501.12948

    2025 arXiv

  7. [15]

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

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

    2025 arXiv

  8. [16]

    Learning to reason at the frontier of learnability

    Thomas Foster and Jakob Foerster. Learning to reason at the frontier of learnability. ArXiv, abs/2502.12272, 2025. https://arxiv.org/abs/2502.12272

    2025 arXiv

  9. [17]

    Training reasoning model with dynamic advantage estimation on reinforcement learning

    Lujun Gui and Qingnan Ren. Training reasoning model with dynamic advantage estimation on reinforcement learning. https://github.com/ShadeCloak/ADORA, 2025

    2025

  10. [18]

    Seed1.5-vl technical report

    Dong Guo, Faming Wu, Feida Zhu, Fuxing Leng, Guang Shi, Haobin Chen, Haoqi Fan, Jian Wang, Jianyu Jiang, Jiawei Wang, Jingji Chen, Jingjia Huang, Kang Lei, Liping Yuan, Lishu Luo, Pengfei Liu, Qinghao Ye, Rui Qian, Shen Yan, Shixiong Zhao, Shuai Peng, Shuangye Li, Sihang Yuan,...

    2025 arXiv

  11. [19]

    Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark

    Yunzhuo Hao, Jiawei Gu, Huichen Will Wang, Linjie Li, Zhengyuan Yang, Lijuan Wang, and Yu Cheng. Can mllms reason in multimodality? emma: An enhanced multimodal reasoning benchmark. ArXiv, abs/2501.05444, 2025. https://arxiv.org/abs/2501.05444

    2025 arXiv

  12. [20]

    Geometry3k dataset

    hiyouga. Geometry3k dataset. https://huggingface.co/datasets/hiyouga/geometry3k, 2025

    2025

  13. [21]

    Praxis-vlm: Vision-grounded decision making via text-driven reinforcement learning

    Zhe Hu, Jing Li, Zhongzhu Pu, Hou Pong Chan, and Yu Yin. Praxis-vlm: Vision-grounded decision making via text-driven reinforcement learning. ArXiv, abs/2503.16965, 2025. https://arxiv.org/abs/2503.16965

    2025

  14. [22]

    From pixels to insights: A survey on automatic chart understanding in the era of large foundation models

    Kung-Hsiang Huang, Hou Pong Chan, May Fung, Haoyi Qiu, Mingyang Zhou, Shafiq Joty, Shih-Fu Chang, and Heng Ji. From pixels to insights: A survey on automatic chart understanding in the era of large foundation models. IEEE Transactions on Knowledge and Data Engineering, 37 0 (5...

    2024

  15. [23]

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

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

    2025 arXiv

  16. [24]

    Dvqa: Understanding data visualizations via question answering

    Kushal Kafle, Brian Price, Scott Cohen, and Christopher Kanan. Dvqa: Understanding data visualizations via question answering. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018. https://openaccess.thecvf.com/content_cvpr_2018/pa...

    2018

  17. [25]

    Figureqa: An annotated figure dataset for visual reasoning

    Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Akos Kadar, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. ArXiv, abs/1710.07300, 2018. https://arxiv.org/abs/1710.07300

    2018 arXiv

  18. [26]

    Geomverse: A systematic evaluation of large models for geometric reasoning

    Mehran Kazemi, Hamidreza Alvari, Ankit Anand, Jialin Wu, Xi Chen, and Radu Soricut. Geomverse: A systematic evaluation of large models for geometric reasoning. In AI for Math Workshop @ ICML 2024, 2024. https://openreview.net/forum?id=1AUbiBrOF1

    2024

  19. [27]

    A diagram is worth a dozen images

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. arXiv, abs/1603.07396, 2016. https://arxiv.org/abs/1603.07396

    2016 arXiv

  20. [28]

    Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension

    Aniruddha Kembhavi, Minjoon Seo, Dustin Schwenk, Jonghyun Choi, Ali Farhadi, and Hannaneh Hajishirzi. Are you smarter than a sixth grader? textbook question answering for multimodal machine comprehension. In Proceedings of the IEEE Conference on Computer Vision and Pattern Rec...

    2017

  21. [30]

    Kimi, Angang Du, Bohong Yin, Bowei Xing, Bowen Qu, Bowen Wang, Cheng Chen, Chenlin Zhang, Chenzhuang Du, Chu Wei, Congcong Wang, Dehao Zhang, Dikang Du, Dongliang Wang, Enming Yuan, Enzhe Lu, Fang Li, Flood Sung, Guangda Wei, Guokun Lai, Han Zhu, Hao Ding, Hao Hu, Hao Yang, Ha...

    2025 arXiv

  22. [31]

    Efficient memory management for large language model serving with pagedattention

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th Symposium on Operating Systems Principles, S...

    2023

  23. [32]

    Mmr1: Advancing the frontiers of multimodal reasoning

    Sicong Leng, Jing Wang, Jiaxi Li, Hao Zhang, Zhiqiang Hu, Boqiang Zhang, Hang Zhang, Yuming Jiang, Xin Li, Deli Zhao, Fan Wang, Yu Rong, Aixin Sun, and Shijian Lu. Mmr1: Advancing the frontiers of multimodal reasoning. https://github.com/LengSicong/MMR1, 2025

    2025

  24. [33]

    Llava-onevision: Easy visual task transfer

    Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. ArXiv, abs/2408.03326, 2024 a . https://arxiv.org/abs/2408.03326

    2024 arXiv

  25. [34]

    Multimodal A r X iv: A dataset for improving scientific comprehension of large vision-language models

    Lei Li, Yuqi Wang, Runxin Xu, Peiyi Wang, Xiachong Feng, Lingpeng Kong, and Qi Liu. Multimodal A r X iv: 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

  26. [35]

    Perception, reason, think, and plan: A survey on large multimodal reasoning models

    Yunxin Li, Zhenyu Liu, Zitao Li, Xuanyu Zhang, Zhenran Xu, Xinyu Chen, Haoyuan Shi, Shenyuan Jiang, Xintong Wang, Jifang Wang, Shouzheng Huang, Xinping Zhao, Borui Jiang, Lanqing Hong, Longyue Wang, Zhuotao Tian, Baoxing Huai, Wenhan Luo, Weihua Luo, Zheng Zhang, Baotian Hu, a...

    2025 arXiv

  27. [36]

    Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L. Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog...

    2023

  28. [37]

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

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

    2022 arXiv

  29. [38]

    Let's learn step by step: Enhancing in-context learning ability with curriculum learning

    Yinpeng Liu, Jiawei Liu, Xiang Shi, Qikai Cheng, Yong Huang, and Wei Lu. Let's learn step by step: Enhancing in-context learning ability with curriculum learning. arXiv preprint arXiv:2402.10738, 2024

    2024 arXiv

  30. [39]

    Othink-mr1: Stimulating multimodal generalized reasoning capabilities via dynamic reinforcement learning

    Zhiyuan Liu, Yuting Zhang, Feng Liu, Changwang Zhang, Ying Sun, and Jun Wang. Othink-mr1: Stimulating multimodal generalized reasoning capabilities via dynamic reinforcement learning. ArXiv, abs/2503.16081, 2025. https://arxiv.org/abs/2503.16081

    2025 arXiv

  31. [40]

    I nter- GPS : Interpretable geometry problem solving with formal language and symbolic reasoning

    Pan Lu, Ran Gong, Shibiao Jiang, Liang Qiu, Siyuan Huang, Xiaodan Liang, and Song-Chun Zhu. I nter- GPS : Interpretable geometry problem solving with formal language and symbolic reasoning. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguist...

    2021

  32. [41]

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

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

    2022

  33. [42]

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

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

    2022

  34. [43]

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

    Pan Lu, Liang Qiu, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, Tanmay Rajpurohit, Peter Clark, and Ashwin Kalyan. Dynamic prompt learning via policy gradient for semi-structured mathematical reasoning. In The Eleventh International Conference on Learning Representations, 2023....

    2023

  35. [44]

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts

    Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. In The Twelfth International Conference on Learning...

    2024

  36. [45]

    Ok-vqa: A visual question answering benchmark requiring external knowledge

    Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2019. https://openaccess.thecvf....

    2019

  37. [46]

    C hart QA : A benchmark for question answering about charts with visual and logical reasoning

    Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. C hart QA : A benchmark for question answering about charts with visual and logical reasoning. In Findings of the Association for Computational Linguistics: ACL 2022, pages 2263--2279, Dublin, Ireland, May...

    2022

  38. [47]

    Minesh Mathew, Viraj Bagal, Rub\`en Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V. Jawahar. Infographicvqa. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1697--1706, January 2022. https://openaccess.thecvf.com/content/W...

    2022

  39. [48]

    Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning

    Fanqing Meng, Lingxiao Du, Zongkai Liu, Zhixiang Zhou, Quanfeng Lu, Daocheng Fu, Tiancheng Han, Botian Shi, Wenhai Wang, Junjun He, et al. Mm-eureka: Exploring the frontiers of multimodal reasoning with rule-based reinforcement learning. ArXiv, abs/2503.07365, 2025. https://ar...

    2025 arXiv

  40. [49]

    Mm-raven dataset

    nimapourjafar. Mm-raven dataset. https://huggingface.co/datasets/nimapourjafar/mm_raven, 2024

    2024

  41. [50]

    Gpt-4o system card

    OpenAI. Gpt-4o system card. ArXiv, abs/2410.21276, 2024 a . https://arxiv.org/abs/2410.21276

    2024 arXiv

  42. [51]

    Openai o1 system card

    OpenAI. Openai o1 system card. ArXiv, abs/2412.16720, 2024 b . https://arxiv.org/abs/2412.16720

    2024 arXiv

  43. [52]

    Introducing o3 and o4-mini, Apr 2025

    OpenAI. Introducing o3 and o4-mini, Apr 2025. https://openai.com/index/introducing-o3-and-o4-mini/

    2025

  44. [53]

    Multimath: Bridging visual and mathematical reasoning for large language models

    Shuai Peng, Di Fu, Liangcai Gao, Xiuqin Zhong, Hongguang Fu, and Zhi Tang. Multimath: Bridging visual and mathematical reasoning for large language models. ArXiv, abs/2409.00147, 2024. https://arxiv.org/abs/2409.00147

    2024 arXiv

  45. [54]

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

    Yi Peng, Peiyu Wang, Xiaokun Wang, Yichen Wei, Jiangbo Pei, Weijie Qiu, Ai Jian, Yunzhuo Hao, Jiachun Pan, Tianyidan Xie, Li Ge, Rongxian Zhuang, Xuchen Song, Yang Liu, and Yahui Zhou. Skywork r1v: Pioneering multimodal reasoning with chain-of-thought. ArXiv, abs/2504.05599, 2...

    2025 arXiv

  46. [55]

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

    Yingzhe Peng, Gongrui Zhang, Miaosen Zhang, Zhiyuan You, Jie Liu, Qipeng Zhu, Kai Yang, Xingzhong Xu, Xin Geng, and Xu Yang. Lmm-r1: Empowering 3b lmms with strong reasoning abilities through two-stage rule-based rl. ArXiv, abs/2503.07536, 2025 b . https://arxiv.org/abs/2503.07536

    2025 arXiv

  47. [56]

    Zero: Memory optimizations toward training trillion parameter models

    Samyam Rajbhandari, Jeff Rasley, Olatunji Ruwase, and Yuxiong He. Zero: Memory optimizations toward training trillion parameter models. ArXiv, abs/1910.02054, 2020. https://arxiv.org/abs/1910.02054

    1910 arXiv

  48. [57]

    Lee, and Sanjeev Arora

    Noam Razin, Zixuan Wang, Hubert Strauss, Stanley Wei, Jason D. Lee, and Sanjeev Arora. What makes a reward model a good teacher? an optimization perspective. ArXiv, abs/2503.15477, 2025. https://arxiv.org/abs/2503.15477

    2025

  49. [58]

    No regrets: Investigating and improving regret approximations for curriculum discovery

    Alex Rutherford, Michael Beukman, Timon Willi, Bruno Lacerda, Nick Hawes, and Jakob Foerster. No regrets: Investigating and improving regret approximations for curriculum discovery. In Advances in Neural Information Processing Systems, volume 37, pages 16071--16101. Curran Ass...

    2024

  50. [59]

    Proximal policy optimization algorithms

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. ArXiv, abs/1707.06347, 2017. https://arxiv.org/abs/1707.06347

    2017 arXiv

  51. [60]

    Solving geometry problems: Combining text and diagram interpretation

    Minjoon Seo, Hannaneh Hajishirzi, Ali Farhadi, Oren Etzioni, and Clint Malcolm. Solving geometry problems: Combining text and diagram interpretation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 1466--1476, Lisbon, Portugal, ...

    2015

  52. [61]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. ArXiv, abs/2402.03300, 2024. https://arxiv.org/abs/2402.03300

    2024 arXiv

  53. [62]

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

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. Vlm-r1: A stable and generalizable r1-style large vision-language model. ArXiv, abs/2504.07615, 2025. https://arxiv...

    2025 arXiv

  54. [63]

    Reason-rft: Reinforcement fine-tuning for visual reasoning

    Huajie Tan, Yuheng Ji, Xiaoshuai Hao, Minglan Lin, Pengwei Wang, Zhongyuan Wang, and Shanghang Zhang. Reason-rft: Reinforcement fine-tuning for visual reasoning. ArXiv, abs/2503.20752, 2025. https://arxiv.org/abs/2503.20752

    2025

  55. [64]

    Kimi Team, Angang Du, Bofei Gao, Bowei Xing, Changjiu Jiang, C Chen, C Li, C Xiao, C Du, C Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms, 2025. URL https://arxiv. org/abs/2501.12599, 2025 a

    2025 arXiv

  56. [65]

    Kwai keye-vl technical report

    Kwai Keye Team, Biao Yang, Bin Wen, Changyi Liu, Chenglong Chu, Chengru Song, Chongling Rao, Chuan Yi, Da Li, Dunju Zang, Fan Yang, Guorui Zhou, Hao Peng, Haojie Ding, Jiaming Huang, Jiangxia Cao, Jiankang Chen, Jingyun Hua, Jin Ouyang, Kaibing Chen, Kaiyu Jiang, Kaiyu Tang, K...

    2025 arXiv

  57. [66]

    Proximal curriculum for reinforcement learning agents

    Georgios Tzannetos, B \'a rbara Gomes Ribeiro, Parameswaran Kamalaruban, and Adish Singla. Proximal curriculum for reinforcement learning agents. Transactions on Machine Learning Research, 2023. https://openreview.net/forum?id=8WUyeeMxMH

    2023

  58. [67]

    Easyarc: Evaluating vision language models on true visual reasoning

    Mert Unsal and Aylin Akkus. Easyarc: Evaluating vision language models on true visual reasoning. arXiv, abs/2506.11595, 2025. https://arxiv.org/abs/2506.11595

    2025 arXiv

  59. [68]

    Vl-rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning

    Haozhe Wang, Chao Qu, Zuming Huang, Wei Chu, Fangzhen Lin, and Wenhu Chen. Vl-rethinker: Incentivizing self-reflection of vision-language models with reinforcement learning. ArXiv, abs/2504.08837, 2025 a . https://arxiv.org/abs/2504.08837

    2025 arXiv

  60. [69]

    Measuring multimodal mathematical reasoning with MATH -vision dataset

    Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with MATH -vision dataset. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 20...

    2024

  61. [70]

    Skywork r1v2: Multimodal hybrid reinforcement learning for reasoning

    Peiyu Wang, Yichen Wei, Yi Peng, Xiaokun Wang, Weijie Qiu, Wei Shen, Tianyidan Xie, Jiangbo Pei, Jianhao Zhang, Yunzhuo Hao, Xuchen Song, Yang Liu, and Yahui Zhou. Skywork r1v2: Multimodal hybrid reinforcement learning for reasoning. ArXiv, abs/2504.16656, 2025 b . https://arx...

    2025 arXiv

  62. [71]

    Visualprm: An effective process reward model for multimodal reasoning

    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. Visualprm: An effective process reward model for multimodal reasoning. ArXiv, abs/2503.10...

    2025 arXiv

  63. [72]

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

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

    2025 arXiv

  64. [73]

    Multimodal chain-of-thought reasoning: A comprehensive survey

    Yaoting Wang, Shengqiong Wu, Yuecheng Zhang, Shuicheng Yan, Ziwei Liu, Jiebo Luo, and Hao Fei. Multimodal chain-of-thought reasoning: A comprehensive survey. ArXiv, abs/2503.12605, 2025 e . https://arxiv.org/abs/2503.12605

    2025 arXiv

  65. [74]

    Dump: Automated distribution-level curriculum learning for rl-based llm post-training

    Zhenting Wang, Guofeng Cui, Yu-Jhe Li, Kun Wan, and Wentian Zhao. Dump: Automated distribution-level curriculum learning for rl-based llm post-training. arXiv preprint arXiv:2504.09710, 2025 f

    2025

  66. [75]

    Visionary-r1: Mitigating shortcuts in visual reasoning with reinforcement learning

    Jiaer Xia, Yuhang Zang, Peng Gao, Yixuan Li, and Kaiyang Zhou. Visionary-r1: Mitigating shortcuts in visual reasoning with reinforcement learning. ArXiv, abs/2505.14677, 2025. https://arxiv.org/abs/2505.14677

    2025

  67. [76]

    Logicvista: Multimodal llm logical reasoning benchmark in visual contexts

    Yijia Xiao, Edward Sun, Tianyu Liu, and Wei Wang. Logicvista: Multimodal llm logical reasoning benchmark in visual contexts. ArXiv, abs/2407.04973, 2024. https://arxiv.org/abs/2407.04973

    2024 arXiv

  68. [77]

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

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

    2025 arXiv

  69. [78]

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization

    Yi Yang, Xiaoxuan He, Hongkun Pan, Xiyan Jiang, Yan Deng, Xingtao Yang, Haoyu Lu, Dacheng Yin, Fengyun Rao, Minfeng Zhu, et al. R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization. ArXiv, abs/2503.10615, 2025. https://arxiv.org/abs/2503.10615

    2025 arXiv

  70. [79]

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

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

    2024 arXiv

  71. [80]

    R1-sharevl: Incentivizing reasoning capability of multimodal large language models via share-grpo

    Huanjin Yao, Qixiang Yin, Jingyi Zhang, Min Yang, Yibo Wang, Wenhao Wu, Fei Su, Li Shen, Minghui Qiu, Dacheng Tao, and Jiaxing Huang. R1-sharevl: Incentivizing reasoning capability of multimodal large language models via share-grpo. ArXiv, abs/2505.16673, 2025. https://arxiv.o...

    2025 arXiv

  72. [81]

    Demystifying long chain-of-thought reasoning in LLM s

    Edward Yeo, Yuxuan Tong, Xinyao Niu, Graham Neubig, and Xiang Yue. Demystifying long chain-of-thought reasoning in LLM s. In ICLR 2025 Workshop on Navigating and Addressing Data Problems for Foundation Models, 2025. https://openreview.net/forum?id=AgtQlhMQ0V

    2025

  73. [82]

    Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi

    Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, Cong Wei, Botao Yu, Ruibin Yuan, Renliang Sun, Ming Yin, Boyuan Zheng, Zhenzhu Yang, Yibo Liu, Wenhao Huang, Huan Sun, Yu Su, and Wenhu Chen. Mmmu: A m...

    2024

  74. [83]

    R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization. ArXiv, abs/2503.12937, 2025. https://arxiv.org/abs/2503.12937

    2025 arXiv

  75. [84]

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

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

    2024

  76. [85]

    Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models

    Jinguo Zhu, Weiyun Wang, Zhe Chen, Zhaoyang Liu, Shenglong Ye, Lixin Gu, Hao Tian, Yuchen Duan, Weijie Su, Jie Shao, Zhangwei Gao, Erfei Cui, Xuehui Wang, Yue Cao, Yangzhou Liu, Xingguang Wei, Hongjie Zhang, Haomin Wang, Weiye Xu, Hao Li, Jiahao Wang, Nianchen Deng, Songze Li,...

    2025 arXiv

Pith tools

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