Pith. sign in

REVIEW 4 major objections 4 minor 47 references

The paper claims that a compact 2B-parameter multimodal model, trained on a new 128k-sample screening dataset with chain-of-thought cold start and a modified GRPO method, surpasses much larger open-source and closed-source models at spottin

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 22:29 UTC pith:SJLRNJLW

load-bearing objection A credible compact-model win on a private image-screening benchmark, but the benchmark's label reliability is unverified and the abstract is out of sync with the full text. the 4 major comments →

arxiv 2511.10055 v2 pith:SJLRNJLW submitted 2025-11-13 cs.CV

Physical Plausibility Reasoning via HCM-GRPO: Empowering Compact Model for Superior Performance

Zhiyuan Hu , Zheng Sun , Yi Wei , Long Yu This is my paper
classification cs.CV
keywords image screeningphysical plausibility reasoningmultimodal large language modelsgroup relative policy optimizationhard cases miningchain-of-thought distillationAI-generated image evaluationpartial-credit reward
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper claims that a 2-billion-parameter multimodal model, trained with chain-of-thought data followed by a modified reinforcement-learning step, can outperform much larger open- and closed-source models at screening AI-generated images for physical plausibility. It introduces a dataset of 128k samples built from real medicine photos, with four objective defect types: appearance deformation, physical shadow, placement layout, and extension rationality. The proposed HCM-GRPO method modifies GRPO in two ways: a partial-credit reward for multi-answer multiple-choice questions, and a hard-case mining stage that re-trains on samples the model got wrong. On the authors' test set the 2B model scores 64.74, while the best existing models cluster near the 34.27 random-guessing baseline. If the result holds, it shows that domain-specific visual screening can be made cheap and reliable without giant models.

Core claim

The central claim is that a compact model can beat large-scale open-source and leading closed-source competitors on physical plausibility image screening when trained with the right two-stage recipe. Stage one uses chain-of-thought data — including a cheap pseudo-labeled split and a smaller human-answer-conditioned split — to give the model a stable reasoning format. Stage two applies HCM-GRPO, which keeps GRPO's group-relative advantage but replaces the binary accuracy reward with a dynamic proportional accuracy reward that gives partial credit when the model's answer is a subset of the correct options, and then refines the model by oversampling the hard cases it previously got wrong. The a

What carries the argument

The load-bearing mechanism is HCM-GRPO, a variant of Group Relative Policy Optimization. GRPO samples several responses, scores each with a reward, and normalises the advantage within the group. HCM-GRPO changes the reward: for a multi-answer question the accuracy reward is len(response)/len(answer) when the response is a subset of the correct answer, giving partial credit instead of a binary 0/1. It also adds hard cases mining: after cold-start SFT, samples that the model answers incorrectly are identified, and in the final training epochs these hard cases are re-added to the full training set so the model is forced to revisit its failures. Together these two changes are what the paper cred

Load-bearing premise

The manual labels for the 1,044 training and 468 test samples are accurate and consistent enough that the reported scores measure physical plausibility reasoning rather than annotation noise.

What would settle it

Re-label the 468 test samples independently with a second team of annotators and measure agreement; if inter-annotator agreement falls below about 0.6 (e.g., Cohen's kappa) on the A/B/C/D/N answers, the 64.74 score and the ranking of models would not be a stable measure of reasoning ability.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A 2B-parameter model reaches 64.74 on the screening test, while leading large and closed-source models score 32–40, so the method's first consequence is a large efficiency gain for the screening task.
  • The DPA reward generalises the idea of partial credit to any multi-answer multiple-choice RL task, and the paper's ablations show it outperforms binary rewards even without hard cases mining.
  • Hard cases mining adds value specifically when combined with the full training set: adding random samples gives +1.07, while adding hard cases gives +4.91 over DPA-GRPO.
  • The method transfers to public benchmarks: HCM-GRPO improves over SFT+GRPO on RealWorldQA, MuirBench, and BLINK, indicating the recipe is not dataset-specific.
  • The exploration split (115,809 unlabeled samples) is left unused; the paper explicitly leaves unsupervised knowledge transfer as future work, so the dataset's full value is untapped.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The manual labels in the 468-sample test set carry the whole comparison; without reported inter-annotator agreement, the 64.74 vs. ~34–40 scores could partly reflect test-set answer ambiguity rather than pure reasoning ability.
  • The dataset is built from medicine photos, so the method's generality to non-medicine products is an open question; re-running HCM-GRPO on a different product category would test whether the gains transfer.
  • The paper's own conclusion says the model still struggles with placement layout and physical shadow; a natural next step is dimension-conditioned CoT data that explicitly contrasts good and bad shadows and layouts.
  • The manuscript switches between 'image aesthetic reasoning' and 'physical plausibility reasoning'; the operational definition here is four objective screening criteria on generated medicine images, so readers should not expect art-critical aesthetics.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes an image screening dataset of over 128k samples (about 640k images) spanning four dimensions — appearance deformation, physical shadow, placement layout, and extension rationality — together with a two-stage training method. Stage 1 is cold-start SFT on CoT data obtained via weakly supervised and answer-driven annotation; Stage 2 is HCM-GRPO, which adds a Dynamic Proportional Accuracy (DPA) reward and hard-case mining to GRPO. The central claim is that HCM-GRPO applied to InternVL3-2B reaches 64.74 on the authors' 468-sample test set, surpassing large open-source MLLMs such as InternVL3-78B and closed-source models such as GPT-4o and Qwen-VL-Max. The paper also reports gains on RealWorldQA, MuirBench, and BLINK.

Significance. If the evaluation is trustworthy, the contribution is meaningful: a large, structured image-screening dataset, a cost-effective CoT annotation pipeline, and a GRPO variant that appears to improve small-model visual reasoning. The ablations are internally consistent and show incremental gains from each component; the public-benchmark results are plausible and support the method's generality. However, the headline claim of surpassing strong closed-source models rests entirely on a private 468-sample test set with no reported inter-annotator agreement or statistical uncertainty. Because the model is trained on the same manual labels used for evaluation, any systematic annotator bias would be learned and rewarded, making the 64.74 score not yet an established result.

major comments (4)
  1. [§3.2, Table 3] The evaluation set is 468 samples labeled by 'manual annotation by human reviewers,' but the paper reports no annotation protocol, number of reviewers, expertise, disagreement rates, or inter-annotator agreement. Since HCM-GRPO is trained on the same manual labels, any systematic bias in the annotators (e.g., over-flagging FLUX artifacts or inconsistent treatment of the 'N' option) would be learned and then rewarded on the test set. This is load-bearing for the claim that HCM-GRPO-2B surpasses GPT-4o and Qwen-VL-Max; without a label-quality audit, the gap may reflect annotation idiosyncrasies rather than general physical-plausibility reasoning. The authors should provide IAA on a subset, describe label adjudication, and ideally release the test set or a random subset for external verification.
  2. [§5.2, Table 3] The reported scores are point estimates on 468 samples with no confidence intervals, significance tests, or multiple seeds. The per-dimension scores are based on even smaller subsamples (e.g., normal images are only 8.5% of the data, so the reported aggregate masks high per-cell variance). For instance, the gap between HCM-GRPO-2B (64.74) and Qwen2.5-VL-72B (39.74) is large, but the paper does not show that the difference is statistically robust. Bootstrap CIs or a per-sample error analysis should be added before the 'superior to closed-source models' claim is accepted.
  3. [§4.2, Eq. (7), vs §5.1 Evaluation Metrics] The DPA reward used during training gives partial credit of len(response)/len(answer) when the response is a subset of the correct answer, while the reported evaluation metric counts a response as correct if it matches any one correct option. These are different objectives: DPA-GRPO optimizes for full-set identification, but the paper evaluates with a relaxed 'any correct option' criterion. The paper should justify this mismatch or evaluate with a metric consistent with the training reward. Without this, the +4.92/+6.20/+11.11 gains from GRPO→DPA-GRPO→HCM-GRPO in Table 6 are not directly comparable to the scores in Table 3.
  4. [Abstract vs. Full Text] The arXiv metadata abstract and the full-text title/abstract disagree in substantive ways. The metadata abstract says 'Physical Plausibility Reasoning' and claims state-of-the-art closed-source models such as GPT5.2 and Gemini3-Pro perform poorly; the full-text title is 'Image Aesthetic Reasoning' and the abstract/Table 3 instead compare against GPT4o and Qwen-VL-Max. GPT5.2 and Gemini3-Pro do not appear anywhere in Table 3 or the full text. This is not a trivial typo: it makes the paper's central benchmarking claim ambiguous and undermines trust in the evaluation's curation. The authors must reconcile the title, abstract, and results tables.
minor comments (4)
  1. [§4.1] The text states 'weakly-supervised CoT data ... with an approximate accuracy of 38.25% (as indicated in Table 3).' Table 3's 38.25 is Qwen-VL-Max's CoT score on the test set, not an accuracy measurement of the pseudo-label CoT data itself. This attribution should be corrected.
  2. [Table 2] Formatting issues: 'T op k', 'T op p', 'T emperature' are rendered as separate words, making the table hard to read.
  3. [§5.4, Tables 8-10] For the public-benchmark experiments, the splits are created by the authors at 1:1 ratio. The paper should clarify whether the split is random, whether the training and test sets contain the same question types, and whether any hyperparameters were tuned on these splits. This is especially important for MuirBench, where SFT already gives +32.87, indicating the validation data is highly in-distribution.
  4. [§2.2] The related-work discussion of VAPO and DAPO is brief but acceptable; consider placing the description of 'hard cases mining' in the related work to distinguish it from curriculum learning or self-training methods.

Circularity Check

0 steps flagged

No significant circularity; one minor non-load-bearing self-citation, plus non-circular data-quality and consistency concerns.

full rationale

Walking the claimed derivation chain (dataset construction -> CoT cold-start -> DPA reward -> hard-case mining -> evaluation), no step reduces to its inputs by construction. The answer-driven CoT data conditions Qwen-VL-Max on the human-annotated answer, but the trained model must produce answers from images at inference; SFT on such data is standard distillation, not a self-fulfilling prediction. The DPA reward and the evaluation metric are aligned but both use the human labels only as ground truth for training/measuring, so the 64.74 test score is an empirical result, not an identity. Hard cases are selected by the SFT model's errors on the training split and oversampled in RL; this is curriculum learning, not a circular definition. The only self-referential element found is a related-work citation to Visual-RFT [7], whose co-author Zheng Sun is also an author here; it is used only to note that RL is applied to object detection and is not load-bearing. Separate non-circular concerns: (i) the abstract claims GPT5.2 and Gemini3-Pro were tested, but Table 3 reports GPT4o and Qwen-VL-Max; (ii) Section 4.1 says weakly-supervised CoT data have 'approximate accuracy of 38.25% (as indicated in Table 3)', but Table 3 is the Qwen-VL-Max CoT score on the test set, not an accuracy measurement of the pseudo-label split; (iii) Section 3.2 asserts 'manual annotation is performed by human reviewers' with no inter-annotator agreement or label-error analysis, a label-reliability risk. These affect the strength of the central comparison but are not circularity. Score 2: at most one minor non-load-bearing self-citation; the central derivation is self-contained and externally checked on RealWorldQA, MuirBench, and BLINK.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 0 invented entities

The central claim rests on standard RL background and on several domain assumptions that are stated but not quantitatively validated: the objectivity of the four annotation dimensions, the correctness of human labels, the usefulness of Qwen-VL-Max-generated CoT data, and the sufficiency of the small test set. Hyperparameters (beta, group size, sampling parameters, epoch schedule) are chosen by hand without sensitivity analysis.

free parameters (5)
  • KL penalty beta = 0.04
    Table 2; chosen by hand for GRPO stability; no sensitivity analysis.
  • Group size G = 4
    Table 2; number of sampled responses per question in GRPO; chosen by hand.
  • Sampling parameters top_k/top_p/temperature = 3 / 0.8 / 1.0
    Table 2; generation stochasticity during RL; chosen by hand.
  • Epoch schedule = 3 general + 2 hard-case epochs
    Section 4.2; the split between general and hard-case training is arbitrary; no ablation on the schedule.
  • Hard-case set = 434 of 1,044 training samples
    Section 5.3/Table 7; determined by the SFT model's errors; the threshold and set size are not analyzed for sensitivity.
axioms (6)
  • domain assumption The four annotated dimensions (appearance deformation, physical shadow, placement layout, extension rationality) are objective and cover physical plausibility.
    Section 3.1/Figure 1; the paper asserts these are straightforward to evaluate objectively, but no inter-annotator agreement study is provided.
  • domain assumption Manual annotations in training and testing splits are ground truth.
    Section 3.2; 'manual annotation is performed by human reviewers to ensure ground-truth accuracy' but no quality metrics or inter-annotator agreement are reported.
  • domain assumption Qwen-VL-Max can generate accurate image descriptions and useful CoT data.
    Section 3.2; the paper states 'we operate under the assumption that the model can generate these basic descriptions with high accuracy'.
  • standard math GRPO framework and its equations are correct and applicable.
    Equations (2)-(5), from DeepSeekMath/DeepSeek-R1; treated as background without re-derivation.
  • domain assumption The 468-sample test set is sufficient to rank models.
    Section 5.1/5.2; no confidence intervals, bootstrap, or multiple seeds are reported.
  • domain assumption The evaluation metric (any correct option counts) matches the intended e-commerce screening use case.
    Section 5.1 'Evaluation Metrics'; the metric is designed for deploying one image on a webpage, but it differs from the DPA reward used in training.

pith-pipeline@v1.3.0-alltime-deepseek · 15827 in / 13250 out tokens · 129401 ms · 2026-08-03T22:29:32.404000+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Physical Plausibility Reasoning via HCM-GRPO: Empowering Compact Model for Superior Performance." pith.science (2026). https://pith.science/paper/SJLRNJLW

@misc{pith2026251110055,
  author       = {Pith},
  title        = {Pith review of: Physical Plausibility Reasoning via HCM-GRPO: Empowering Compact Model for Superior Performance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJLRNJLW}},
  note         = {Machine review of arXiv:2511.10055}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The performance of image generation has been significantly improved in recent years. However, the study of image screening is rare, and its performance with Multimodal Large Language Models (MLLMs) is unsatisfactory due to the lack of data and the weak physical plausibility reasoning ability in MLLMs. In this work, we propose a complete solution to address these problems in terms of data and methodology. For data, we collect a comprehensive image screening dataset with over 128k samples, comprising about 640k images. Each sample consists of an original image and four generated images. The dataset evaluates the physical plausibility reasoning ability under four aspects: appearance deformation, physical shadow, placement layout, and extension rationality. Regarding data annotation, we investigate multiple approaches, including purely manual, fully automated, and answer-driven annotations, to acquire high-quality chains of thought (CoT) data in the most cost-effective manner. Methodologically, we introduce a Hard Cases Mining (HCM) strategy with a Dynamic Proportional Accuracy (DPA) reward into the Group Relative Policy Optimization (GRPO) framework, called HCM-GRPO. This enhanced method demonstrates superior physical plausibility reasoning capabilities compared to the original GRPO. Our experimental results reveal that even state-of-the-art closed-source MLLMs, such as GPT5.2 and Gemini3-Pro, exhibit unsatisfactory performance in physical plausibility reasoning. In contrast, by leveraging the HCM-GRPO, we are able to surpass the scores of both large-scale open-source and leading closed-source models with a much smaller model.

Figures

Figures reproduced from arXiv: 2511.10055 by Long Yu, Yi Wei, Zheng Sun, Zhiyuan Hu.

Figure 1
Figure 1. Figure 1: Overview of the image aesthetic dataset and quantitative comparison results. (a) We summarize [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of dataset construction pipeline. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Presentation of different annotation paradigms. 3.2. Dataset Division and Annotation As shown in [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Illustration of model training process. each reward and the mean of the reward score group, which can be regarded as the baseline value, and normalizes the differences using the standard deviation: Ai = ri − mean({r1,r2, ...,rG}) std({r1,r2, ...,rG}) , (2) where Ai quantifies the relative quality of the i-th response in comparison to other candidates within the same sampled group. Based on the simple advan… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 27 linked inside Pith

  1. [1]

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, Y . Duan, H. Tian, W. Su, J. Shao, et al., Internvl3: Exploring advanced training and test-time recipes for open-source multimodal models, arXiv preprint arXiv:2504.10479 (2025)

    Pith/arXiv arXiv 2025

  2. [2]

    S. Bai, K. Chen, X. Liu, J. Wang, W. Ge, S. Song, K. Dang, P. Wang, S. Wang, J. Tang, et al., Qwen2. 5-vl technical report, arXiv preprint arXiv:2502.13923 (2025)

    Pith/arXiv arXiv 2025

  3. [3]

    S. Lu, Y . Li, Q.-G. Chen, Z. Xu, W. Luo, K. Zhang, H.-J. Ye, Ovis: Struc- tural embedding alignment for multimodal large language model, arXiv preprint arXiv:2405.20797 (2024)

    Pith/arXiv arXiv 2024

  4. [4]

    H. Cai, Y . Yang, W. Hu, Mm-iq: Benchmarking human-like abstraction and rea- soning in multimodal models, arXiv preprint arXiv:2502.00698 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Z. Yang, J. Tang, Z. Li, P. Wang, J. Wan, H. Zhong, X. Liu, M. Yang, P. Wang, Y . Liu, et al., Cc-ocr: A comprehensive and challenging ocr benchmark for evaluating large multimodal models in literacy, arXiv preprint arXiv:2412.02210 (2024)

    Pith/arXiv arXiv 2024

  6. [6]

    Cheng, Q

    Z. Cheng, Q. Chen, J. Zhang, H. Fei, X. Feng, W. Che, M. Li, L. Qin, Comt: A novel benchmark for chain of multi-modal thought on large vision-language mod- els, in: Proceedings of the AAAI Conference on Artificial Intelligence, V ol. 39, 2025, pp. 23678–23686

    2025

  7. [7]

    Z. Liu, Z. Sun, Y . Zang, X. Dong, Y . Cao, H. Duan, D. Lin, J. Wang, Visual-rft: Visual reinforcement fine-tuning, arXiv preprint arXiv:2503.01785 (2025)

    Pith/arXiv arXiv 2025

  8. [8]

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

    Pith/arXiv arXiv 2025

  9. [9]

    H. Wang, C. Qu, Z. Huang, W. Chu, F. Lin, W. Chen, Vl-rethinker: Incentiviz- ing self-reflection of vision-language models with reinforcement learning, arXiv preprint arXiv:2504.08837 (2025). 29

    Pith/arXiv arXiv 2025

  10. [10]

    X. Chen, Z. Wu, X. Liu, Z. Pan, W. Liu, Z. Xie, X. Yu, C. Ruan, Janus-pro: Unified multimodal understanding and generation with data and model scaling, arXiv preprint arXiv:2501.17811 (2025)

    Pith/arXiv arXiv 2025

  11. [11]

    Q. Sun, Y . Cui, X. Zhang, F. Zhang, Q. Yu, Y . Wang, Y . Rao, J. Liu, T. Huang, X. Wang, Generative multimodal models are in-context learners, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14398–14409

    2024

  12. [12]

    K. Tian, Y . Jiang, Z. Yuan, B. Peng, L. Wang, Visual autoregressive modeling: Scalable image generation via next-scale prediction, Advances in neural informa- tion processing systems 37 (2024) 84839–84865

    2024

  13. [13]

    Huang, S

    Z. Huang, S. Zhuang, C. Fu, B. Yang, Y . Zhang, C. Sun, Z. Zhang, Y . Wang, C. Li, Z.-J. Zha, Wegen: A unified model for interactive multimodal generation as we chat, arXiv preprint arXiv:2503.01115 (2025)

    Pith/arXiv arXiv 2025

  14. [14]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33 (2020) 6840–6851

    2020

  15. [15]

    Rombach, A

    R. Rombach, A. Blattmann, D. Lorenz, P. Esser, B. Ommer, High-resolution im- age synthesis with latent diffusion models, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10684–10695

    2022

  16. [16]

    Podell, Z

    D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, R. Rombach, Sdxl: Improving latent diffusion models for high-resolution image synthesis, arXiv preprint arXiv:2307.01952 (2023)

    Pith/arXiv arXiv 2023

  17. [17]

    Peebles, S

    W. Peebles, S. Xie, Scalable diffusion models with transformers, in: Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 4195– 4205

    2023

  18. [18]

    H. Ye, J. Zhang, S. Liu, X. Han, W. Yang, Ip-adapter: Text compati- ble image prompt adapter for text-to-image diffusion models, arXiv preprint arXiv:2308.06721 (2023). 30

    Pith/arXiv arXiv 2023

  19. [19]

    Zhang, Z

    H. Zhang, Z. Duan, X. Wang, Y . Chen, Y . Zhao, Y . Zhang, Nexus-gen: A unified model for image understanding, generation, and editing, arXiv preprint arXiv:2504.21356 (2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Zhang, A

    L. Zhang, A. Rao, M. Agrawala, Adding conditional control to text-to-image diffusion models, in: Proceedings of the IEEE/CVF international conference on computer vision, 2023, pp. 3836–3847

    2023

  21. [21]

    Z. Tan, S. Liu, X. Yang, Q. Xue, X. Wang, Ominicontrol: Minimal and universal control for diffusion transformer, arXiv preprint arXiv:2411.15098 (2024)

    Pith/arXiv arXiv 2024

  22. [22]

    H. Zhu, Y . Zhou, R. Yao, G. Wang, Y . Yang, Learning image aesthetic subjectivity from attribute-aware relational reasoning network, Pattern Recognition Letters 155 (2022) 84–91

    2022

  23. [23]

    L. Li, Y . Huang, J. Wu, Y . Yang, Y . Li, Y . Guo, G. Shi, Theme-aware visual at- tribute reasoning for image aesthetics assessment, IEEE Transactions on Circuits and Systems for Video Technology 33 (9) (2023) 4798–4811

    2023

  24. [24]

    Z. Zhou, Q. Wang, B. Lin, Y . Su, R. Chen, X. Tao, A. Zheng, L. Yuan, P. Wan, D. Zhang, Uniaa: A unified multi-modal image aesthetic assessment baseline and benchmark, arXiv preprint arXiv:2404.09619 (2024)

    Pith/arXiv arXiv 2024

  25. [25]

    Zheng, D

    P. Zheng, D. Gao, D.-P. Fan, L. Liu, J. Laaksonen, W. Ouyang, N. Sebe, Bilateral reference for high-resolution dichotomous image segmentation, CAAI Artificial Intelligence Research (2024)

    2024

  26. [26]

    B. F. Labs, Flux,https://github.com/black-forest-labs/flux(2024)

    2024

  27. [27]

    Z. Shao, P. Wang, Q. Zhu, R. Xu, J. Song, X. Bi, H. Zhang, M. Zhang, Y . Li, Y . Wu, et al., Deepseekmath: Pushing the limits of mathematical reasoning in open language models, arXiv preprint arXiv:2402.03300 (2024)

    Pith/arXiv arXiv 2024

  28. [28]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al., Deepseek-r1: Incentivizing reasoning capability in llms via rein- forcement learning, arXiv preprint arXiv:2501.12948 (2025). 31

    Pith/arXiv arXiv 2025

  29. [29]

    Z. Chen, W. Wang, H. Tian, S. Ye, Z. Gao, E. Cui, W. Tong, K. Hu, J. Luo, Z. Ma, et al., How far are we to gpt-4v? closing the gap to commercial multimodal mod- els with open-source suites, Science China Information Sciences 67 (12) (2024) 220101

    2024

  30. [30]

    W. Wang, Z. Chen, W. Wang, Y . Cao, Y . Liu, Z. Gao, J. Zhu, X. Zhu, L. Lu, Y . Qiao, et al., Enhancing the reasoning ability of multimodal large language models via mixed preference optimization, arXiv preprint arXiv:2411.10442 (2024)

    Pith/arXiv arXiv 2024

  31. [31]

    W. Hong, W. Yu, X. Gu, G. Wang, G. Gan, H. Tang, J. Cheng, J. Qi, J. Ji, L. Pan, et al., Glm-4.1 v-thinking: Towards versatile multimodal reasoning with scalable reinforcement learning, arXiv e-prints (2025) arXiv–2507

    2025

  32. [32]

    B. Li, Z. Lin, W. Peng, J. d. D. Nyandwi, D. Jiang, Z. Ma, S. Khanuja, R. Krishna, G. Neubig, D. Ramanan, Naturalbench: Evaluating vision-language models on natural adversarial samples, arXiv preprint arXiv:2410.14669 (2024)

    Pith/arXiv arXiv 2024

  33. [33]

    C.-H. Yeh, C. Wang, S. Tong, T.-Y . Cheng, R. Wang, T. Chu, Y . Zhai, Y . Chen, S. Gao, Y . Ma, Seeing from another perspective: Evaluating multi-view under- standing in mllms, arXiv preprint arXiv:2504.15280 (2025)

    Pith/arXiv arXiv 2025

  34. [34]

    Q. Yu, Z. Zhang, R. Zhu, Y . Yuan, X. Zuo, Y . Yue, T. Fan, G. Liu, L. Liu, X. Liu, et al., Dapo: An open-source llm reinforcement learning system at scale, 2025, URL https://arxiv. org/abs/2503.14476

    Pith/arXiv arXiv 2025

  35. [35]

    Y . Yuan, Q. Yu, X. Zuo, R. Zhu, W. Xu, J. Chen, C. Wang, T. Fan, Z. Du, X. Wei, et al., Vapo: Efficient and reliable reinforcement learning for advanced reasoning tasks, arXiv preprint arXiv:2504.05118 (2025)

    Pith/arXiv arXiv 2025

  36. [36]

    Zhang, Y

    K. Zhang, Y . Hong, J. Bao, H. Jiang, Y . Song, D. Hong, H. Xiong, Gvpo: Group variance policy optimization for large language model post-training, arXiv preprint arXiv:2504.19599 (2025). 32

    arXiv 2025

  37. [37]

    Zheng, S

    C. Zheng, S. Liu, M. Li, X.-H. Chen, B. Yu, C. Gao, K. Dang, Y . Liu, R. Men, A. Yang, et al., Group sequence policy optimization, arXiv preprint arXiv:2507.18071 (2025)

    Pith/arXiv arXiv 2025

  38. [38]

    Y . Wang, Z. Li, Y . Zang, C. Wang, Q. Lu, C. Jin, J. Wang, Unified multimodal chain-of-thought reward model through reinforcement fine-tuning, arXiv preprint arXiv:2505.03318 (2025)

    arXiv 2025

  39. [39]

    Zhang, J

    J. Zhang, J. Huang, H. Yao, S. Liu, X. Zhang, S. Lu, D. Tao, R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization, arXiv preprint arXiv:2503.12937 (2025)

    Pith/arXiv arXiv 2025

  40. [40]

    Alibaba Cloud, Qwen-vl-max - alibaba cloud,https://www.aliyun.com

  41. [41]

    W. Wang, Z. Gao, L. Gu, H. Pu, L. Cui, X. Wei, Z. Liu, L. Jing, S. Ye, J. Shao, et al., Internvl3. 5: Advancing open-source multimodal models in versatility, rea- soning, and efficiency, arXiv preprint arXiv:2508.18265 (2025)

    Pith/arXiv arXiv 2025

  42. [42]

    OpenAI, Gpt-4o - openai,https://openai.com

  43. [43]

    Alibaba Cloud, Qwen-vl-plus - alibaba cloud,https://www.aliyun.com

  44. [44]

    Y . Zhao, J. Huang, J. Hu, X. Wang, Y . Mao, D. Zhang, Z. Jiang, Z. Wu, B. Ai, A. Wang, W. Zhou, Y . Chen, Swift:a scalable lightweight infrastructure for fine- tuning (2024). arXiv:2408.05517. URLhttps://arxiv.org/abs/2408.05517

    Pith/arXiv arXiv 2024

  45. [45]

    X.AI Corp, Grok-1.5 vision preview: Connecting the digital and physical worlds with our first multimodal model,https://x.ai/blog/grok-1.5v

  46. [46]

    F. Wang, X. Fu, J. Y . Huang, Z. Li, Q. Liu, X. Liu, M. D. Ma, N. Xu, W. Zhou, K. Zhang, et al., Muirbench: A comprehensive benchmark for robust multi-image understanding, arXiv preprint arXiv:2406.09411 (2024)

    Pith/arXiv arXiv 2024

  47. [47]

    X. Fu, Y . Hu, B. Li, Y . Feng, H. Wang, X. Lin, D. Roth, N. A. Smith, W.-C. Ma, R. Krishna, Blink: Multimodal large language models can see but not perceive, in: European Conference on Computer Vision, Springer, 2024, pp. 148–166. 33

    2024