Pith. sign in

REVIEW 3 major objections 74 references

Multimodal reasoning gets better when teachers switch by token, not by whole sample.

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 · grok-4.5

2026-07-12 09:28 UTC pith:4LKICJW4

load-bearing objection Solid multimodal OPD systems paper with real gains, but the headline mechanism (entropy token arbitration) is not isolated from multi-teacher ensembling plus description transfer. the 3 major comments →

arxiv 2607.02592 v1 pith:4LKICJW4 submitted 2026-07-01 cs.CV cs.LG

H-OPD: Confidence Aware Heterogeneous Multi-Teacher Multimodal On-policy Distillation

Qixiang Yin , Huanjin Yao , Yuchen Cai , Jianghao Chen , Ziyi Wang , Min Yang , Fei Su , Zhicheng Zhao This is my paper
classification cs.CV cs.LG
keywords on-policy distillationmultimodal reasoningmulti-teacher distillationtoken-level arbitrationvision-language modelsknowledge distillationconfidence-aware fusion
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.

This paper argues that on-policy distillation for multimodal models fails when each sample is locked to one teacher for the whole answer. Visual grounding and abstract reasoning matter at different decoding steps, so a single teacher is a poor guide for the full trajectory. H-OPD instead runs a vision-language teacher and a text-only teacher on the same student-generated path, hands the text teacher a written proxy of the image, and at every token picks or mixes them by confidence. Across eleven reasoning benchmarks the method beats standard on-policy distillation and group-relative policy optimization, with faster gains per training token. The practical claim is that complementary teachers should share a trajectory and be arbitrated step by step, not assigned once per problem.

Core claim

The authors claim that heterogeneous teachers are complementary inside one multimodal reasoning trajectory, and that replacing sample- or task-level teacher routing with confidence-aware token-level arbitration along the student's own rollouts produces stronger student reasoning than single-teacher on-policy distillation or GRPO.

What carries the argument

Confidence-aware token-level teacher arbitration: at each student decoding step, predictive entropy on each teacher's truncated top-k distribution sets a soft weight that fuses the vision-language teacher and the text-only teacher into one target distribution for reverse KL distillation; vision-to-language description transfer first writes key image content into text so the text teacher can participate.

Load-bearing premise

Lower predictive entropy on a teacher's truncated next-token distribution is treated as a trustworthy signal that this teacher should dominate that token.

What would settle it

Ablate entropy-based weights against uniform averaging or random token assignment of the same two teachers on the same student trajectories; if the reported gains over single-teacher OPD disappear, the arbitration claim fails. Separately, measure how often the lower-entropy teacher is wrong on held-out steps—if low entropy systematically misranks correctness, the proxy collapses.

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

If this is right

  • Post-training for multimodal reasoners should expose each student trajectory to both vision-language and text-only teachers rather than routing whole samples to one modality specialist.
  • Token-level confidence fusion can raise average accuracy over single-teacher OPD and GRPO on diverse math, chart, logic, and VQA suites.
  • The same recipe continues to help when both the student and the teacher pair are scaled up, and when the vision-language teacher is itself strengthened by RL.
  • On-policy distillation under this scheme can reach higher accuracy with fewer generated training tokens than sequence-level RL methods like GRPO.

Where Pith is reading between the lines

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

  • If token-state reliability really swings within one answer, any fixed mixture of experts (not only VL vs text) may be leaving gains on the table until arbitration is moved to the token level.
  • Description quality is a hidden bottleneck: better or jointly trained visual proxies could raise the text teacher's ceiling without changing the arbitration math.
  • Entropy arbitration may need calibration or correctness-aware signals when both teachers are confidently wrong, a failure mode the method already flags but does not solve.

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

3 major / 0 minor

Summary. The paper proposes H-OPD, a heterogeneous multi-teacher on-policy distillation framework for multimodal reasoning. Instead of routing each sample or task to a single teacher, it exposes the same student trajectory to a vision-language (VL) teacher and a text-only teacher, bridges the modality gap via offline vision-to-language description transfer (Eq. 5), and fuses the teachers at each token with a confidence-aware weight derived from negative entropy over renormalized top-k distributions (Eqs. 7–9). The student is trained by reverse KL on the arbitrated target restricted to the merged top-k support (Eq. 10). Empirically, on Qwen3-VL 2B/4B students and 4B/8B teacher pairs, H-OPD improves average accuracy over GRPO, SFT, single-teacher OPD, and ExOPD across 11 multimodal reasoning benchmarks (Table 1), with additional evidence from a stronger RL teacher (Table 2), τf/top-k ablations (Table 3), token-efficiency curves (Fig. 5), a complementarity pilot (Fig. 3), and a qualitative λt heatmap (Fig. 6).

Significance. If the gains are genuinely attributable to token-level heterogeneous arbitration rather than multi-teacher ensembling or description-augmented text supervision alone, the work is a clear and useful advance for multimodal post-training: it reframes teacher routing from sample/task level to state-dependent token level, provides a practical modality bridge for text-only teachers, and reports consistent lifts on standard public benchmarks with multi-scale student/teacher settings and better token efficiency than GRPO. The pilot diagnostics, restricted-support distillation design, and public-code commitment are strengths. The contribution is primarily empirical/systems rather than theoretical; its lasting value depends on isolating the arbitration mechanism from simpler multi-teacher baselines.

major comments (3)
  1. The central claim attributes Table 1 gains (e.g., 2B student avg 51.1 OPD → 53.5 H-OPD; larger jumps on LogicVista/CharXiv) to confidence-aware token-level arbitration (Sec. 3.4, Eqs. 7–9). Main comparisons are only against single-VL OPD/ExOPD/GRPO/SFT. Missing load-bearing controls with the same two teachers and the same offline descriptions: (i) uniform average λt=0.5, (ii) hard max-confidence selection (τf→0), (iii) fixed sample-level mixture α(x) as in Eq. 2, and (iv) text-teacher-only with description transfer. Without these, the lift is consistent with multi-teacher ensembling plus GPT-filtered description transfer rather than with token-level arbitration specifically. These ablations are necessary to support the paper’s main mechanistic claim.
  2. Sec. 3.2 / Fig. 3 establish that VL and text teachers disagree (Jaccard not near 1) and that entropy differences flip sign, but they do not show that entropy-weighted λt improves student accuracy over fixed mixtures or hard selection. Sec. 3.4 treats negative entropy of truncated top-k distributions as the reliability proxy (Eqs. 7–8); the Limitations section already notes that low entropy need not imply token-level correctness under overconfidence/miscalibration. A direct validation—e.g., correlation of λt with token correctness on held-out rollouts, or accuracy under entropy vs. alternative confidence scores (max-prob, margin)—is needed for the arbitration design to be more than a plausible heuristic.
  3. The text teacher’s contribution depends on offline VL-generated descriptions filtered by GPT-4.1-mini (Sec. 3.3, Sec. 4.1). There is no ablation of description quality (unfiltered vs. filtered; VL-teacher captions vs. stronger captioner; noisy/incomplete descriptions). Given that Limitations flags this dependence, and that description transfer is a free parameter of the method, at least one controlled comparison is required to bound how much of the Table 1 gain is description quality versus arbitration.

Circularity Check

0 steps flagged

No significant circularity: empirical multi-teacher OPD evaluated on external public benchmarks; results are not forced by definition or self-citation.

full rationale

H-OPD is an empirical systems paper. Its load-bearing chain is (i) a pilot showing VL and text teachers disagree and that entropy differences flip sign along trajectories (Fig. 3), (ii) a concrete training objective—reverse KL of the student to a confidence-weighted mixture of truncated teacher distributions on student rollouts (Eqs. 7–10), enabled by offline vision-to-language description transfer (Eq. 5)—and (iii) evaluation on 11 external public benchmarks (MathVista, MathVision, CharXiv, LogicVista, etc.) against GRPO/SFT/OPD/ExOPD (Table 1). None of these steps redefine the reported metric in terms of a fitted constant, import a uniqueness theorem from the authors, or rename a known result as a derivation. Self-citations (e.g., Yao/Cai related work) appear only as ordinary background and are not load-bearing for the central claim. Missing controls (uniform average vs entropy arbitration) are an experimental-design concern, not circularity. Score 0 is therefore appropriate.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 1 invented entities

Load-bearing content is mostly standard OPD math plus empirical design choices. Free parameters are training/arbitration hyperparameters chosen by ablation. Domain assumptions include complementarity of VL vs text teachers and entropy-as-confidence. The main invented construct is the confidence-aware arbitrated target on the merged top-k support, validated only inside this paper’s experiments.

free parameters (3)
  • fusion temperature τf = 0.5 (best in ablation)
    Controls sharpness of teacher mixing (Eq. 8); ablated in Table 3a with best reported at 0.5 (main text also mentions 1.0 in implementation details).
  • top-k candidate size = 8
    Truncates each teacher’s support before merge (Eq. 6); ablated in Table 3b; k=8 chosen as efficiency/performance trade-off.
  • student learning rate / batch size = 1e-6 / 128
    Adam LR 1e-6 and batch 128 set for training; not derived, chosen for optimization stability.
axioms (4)
  • domain assumption On-policy reverse KL distillation on student-generated trajectories is a valid post-training objective for aligning multimodal students (Sec. 3.1, Eq. 1).
    Inherited from prior OPD literature; not re-proved here.
  • domain assumption VL and text-only teachers provide complementary token-level supervision within the same multimodal trajectory (Sec. 3.2; pilot Fig. 3).
    Motivated by cited perception-vs-reasoning failure modes and supported by a small pilot, not a theorem.
  • ad hoc to paper Negative entropy of renormalized top-k teacher distributions is a usable confidence score for arbitration (Sec. 3.4, Eq. 7–8).
    Central design choice; Limitations notes calibration failures can break the proxy.
  • ad hoc to paper Offline VL-generated image descriptions sufficiently transfer task-relevant visual semantics to a text-only teacher (Sec. 3.3, Eq. 5).
    Required for heterogeneous supervision; quality filtering with GPT-4.1-mini is used but residual noise remains a stated limitation.
invented entities (1)
  • Confidence-aware arbitrated target distribution q_A^(t) on merged top-k support no independent evidence
    purpose: Provide a single token-level teacher mixture that dynamically weights VL vs text supervision for the student KL loss.
    Defined in Eqs. 6–9 as the paper’s core methodological object; evidence is internal empirical gains only.

pith-pipeline@v1.1.0-grok45 · 24577 in / 3295 out tokens · 44700 ms · 2026-07-12T09:28:07.935591+00:00 · methodology

0 comments
read the original abstract

On-policy distillation (OPD) has recently emerged as an effective post-training paradigm by providing supervision on student-generated trajectories. However, existing OPD methods for multimodal reasoning usually rely on a static teacher routing, assigning each sample to a single teacher based on modality or task type. This ignores that visual grounding and abstract reasoning may dominate different decoding steps, making a single teacher insufficient for the full trajectory. To this end, H-OPD is proposed as a confidence-aware heterogeneous multi-teacher OPD framework for multimodal reasoning. By verifying the complementarity of heterogeneous teachers in the same reasoning process, H-OPD replaces task or sample level teacher routing with token-level teacher arbitration along the shared student trajectory. H-OPD employs vision-to-language description transfer to enable text-only teachers to access key visual semantics, and uses a confidence-aware arbitration mechanism to dynamically combine vision-language teacher and text-only teachers at each token. Extensive evaluations over 11 widely-used reasoning benchmarks showcase the superior performance of our method.

Figures

Figures reproduced from arXiv: 2607.02592 by Fei Su, Huanjin Yao, Jianghao Chen, Min Yang, Qixiang Yin, Yuchen Cai, Zhicheng Zhao, Ziyi Wang.

Figure 1
Figure 1. Figure 1: Performance comparison across 11 multi￾modal reasoning benchmarks. H-OPD consistently out￾performs standard OPD and GRPO, leading to stronger overall performance. student-generated trajectories and has emerged as a promising paradigm for reasoning post-training. However, an effective framework for high-quality, fine-grained, and token-state-aware supervision is still lacking in multimodal reasoning. As ill… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison between conventional task-level routing and our token-level arbitration OPD framework. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Pilot experiment on heterogeneous teacher [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Overview of H-OPD, which consists of a vision-to-language description transfer module for enabling [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of training token efficiency be [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of the token-level teacher arbitration in H-OPD. (a) Input image. (b) Input question. (c) [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Case studies [PITH_FULL_IMAGE:figures/full_fig_p016_7.png] 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

74 extracted references · 39 linked inside Pith

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    doi:10.1145/322234.322243 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    5-vl technical report , author=

    Qwen2. 5-vl technical report , author=. arXiv preprint arXiv:2502.13923 , year=

    Pith/arXiv arXiv

  9. [9]

    Proceedings of the 32nd ACM International Conference on Multimedia , pages=

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models , author=. Proceedings of the 32nd ACM International Conference on Multimedia , pages=

  10. [10]

    arXiv preprint arXiv:2511.21631 , year=

    Qwen3-VL Technical Report , author=. arXiv preprint arXiv:2511.21631 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2503.12937 , year=

    R1-vl: Learning to reason with multimodal large language models via step-wise group relative policy optimization , author=. arXiv preprint arXiv:2503.12937 , year=

    Pith/arXiv arXiv

  12. [12]

    arXiv preprint arXiv:2503.10615 , year=

    R1-onevision: Advancing generalized multimodal reasoning through cross-modal formalization , author=. arXiv preprint arXiv:2503.10615 , year=

    Pith/arXiv arXiv

  13. [13]

    arXiv preprint arXiv:2505.16673 , year=

    R1-ShareVL: Incentivizing Reasoning Capability of Multimodal Large Language Models via Share-GRPO , author=. arXiv preprint arXiv:2505.16673 , year=

    Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2511.16672 , year=

    EvoLMM: Self-Evolving Large Multimodal Models with Continuous Rewards , author=. arXiv preprint arXiv:2511.16672 , year=

    Pith/arXiv arXiv

  15. [15]

    arXiv preprint arXiv:2511.15661 , year=

    VisPlay: Self-Evolving Vision-Language Models from Images , author=. arXiv preprint arXiv:2511.15661 , year=

    arXiv

  16. [16]

    arXiv preprint arXiv:2509.25541 , year=

    Vision-zero: Scalable vlm self-improvement via strategic gamified self-play , author=. arXiv preprint arXiv:2509.25541 , year=

    arXiv

  17. [17]

    arXiv preprint arXiv:2511.19900 , year=

    Agent0-VL: Exploring Self-Evolving Agent for Tool-Integrated Vision-Language Reasoning , author=. arXiv preprint arXiv:2511.19900 , year=

    arXiv

  18. [18]

    arXiv preprint arXiv:2505.03335 , year=

    Absolute zero: Reinforced self-play reasoning with zero data , author=. arXiv preprint arXiv:2505.03335 , year=

    Pith/arXiv arXiv

  19. [19]

    arXiv preprint arXiv:2508.05004 , year=

    R-zero: Self-evolving reasoning llm from zero data , author=. arXiv preprint arXiv:2508.05004 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2504.16084 , year=

    Ttrl: Test-time reinforcement learning , author=. arXiv preprint arXiv:2504.16084 , year=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2505.22453 , year=

    Unsupervised Post-Training for Multi-Modal LLM Reasoning via GRPO , author=. arXiv preprint arXiv:2505.22453 , year=

    arXiv

  22. [22]

    arXiv preprint arXiv:2510.24285 , year=

    Viper: Empowering the self-evolution of visual perception abilities in vision-language model , author=. arXiv preprint arXiv:2510.24285 , year=

    arXiv

  23. [23]

    arXiv preprint arXiv:2505.20347 , year=

    SeRL: Self-Play Reinforcement Learning for Large Language Models with Limited Data , author=. arXiv preprint arXiv:2505.20347 , year=

    arXiv

  24. [24]

    arXiv preprint arXiv:2506.08745 , year=

    Consistent Paths Lead to Truth: Self-Rewarding Reinforcement Learning for LLM Reasoning , author=. arXiv preprint arXiv:2506.08745 , year=

    Pith/arXiv arXiv

  25. [25]

    arXiv preprint arXiv:2508.11356 , year=

    Ettrl: Balancing exploration and exploitation in llm test-time reinforcement learning via entropy mechanism , author=. arXiv preprint arXiv:2508.11356 , year=

    Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2505.22660 , year=

    Maximizing Confidence Alone Improves Reasoning , author=. arXiv preprint arXiv:2505.22660 , year=

    Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:2502.17419 , year=

    From system 1 to system 2: A survey of reasoning large language models , author=. arXiv preprint arXiv:2502.17419 , year=

    Pith/arXiv arXiv

  28. [28]

    arXiv preprint arXiv:2310.02255 , year=

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts , author=. arXiv preprint arXiv:2310.02255 , year=

    Pith/arXiv arXiv

  29. [29]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  30. [30]

    Advances in neural information processing systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

  31. [31]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Rlhf-v: Towards trustworthy mllms via behavior alignment from fine-grained correctional human feedback , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  32. [32]

    arXiv preprint arXiv:2412.18319 , year=

    Mulberry: Empowering mllm with o1-like reasoning and reflection via collective monte carlo tree search , author=. arXiv preprint arXiv:2412.18319 , year=

    Pith/arXiv arXiv

  33. [33]

    rlhf: Scaling reinforcement learning from human feedback with ai feedback , author=

    Rlaif vs. rlhf: Scaling reinforcement learning from human feedback with ai feedback , author=. arXiv preprint arXiv:2309.00267 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    Pith/arXiv arXiv

  35. [35]

    arXiv preprint arXiv:2503.06749 , year=

    Vision-r1: Incentivizing reasoning capability in multimodal large language models , author=. arXiv preprint arXiv:2503.06749 , year=

    Pith/arXiv arXiv

  36. [36]

    arXiv preprint arXiv:2503.06520 , year=

    Seg-zero: Reasoning-chain guided segmentation via cognitive reinforcement , author=. arXiv preprint arXiv:2503.06520 , year=

    Pith/arXiv arXiv

  37. [37]

    arXiv preprint arXiv:2507.03019 , year=

    Look-back: Implicit visual re-focusing in mllm reasoning , author=. arXiv preprint arXiv:2507.03019 , year=

    Pith/arXiv arXiv

  38. [38]

    arXiv preprint arXiv:2402.03300 , year=

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models , author=. arXiv preprint arXiv:2402.03300 , year=

    Pith/arXiv arXiv

  39. [39]

    arXiv preprint arXiv:2602.12125 , year=

    Learning beyond teacher: Generalized on-policy distillation with reward extrapolation , author=. arXiv preprint arXiv:2602.12125 , year=

    Pith/arXiv arXiv

  40. [40]

    Thinking Machines Lab: Connectionism , year =

    Kevin Lu and Thinking Machines Lab , title =. Thinking Machines Lab: Connectionism , year =

  41. [41]

    arXiv preprint arXiv:2604.00626 , year=

    A Survey of On-Policy Distillation for Large Language Models , author=. arXiv preprint arXiv:2604.00626 , year=

    Pith/arXiv arXiv

  42. [42]

    The twelfth international conference on learning representations , year=

    Minillm: Knowledge distillation of large language models , author=. The twelfth international conference on learning representations , year=

  43. [43]

    The twelfth international conference on learning representations , year=

    On-policy distillation of language models: Learning from self-generated mistakes , author=. The twelfth international conference on learning representations , year=

  44. [44]

    arXiv preprint arXiv:2604.16830 , year=

    The Illusion of Certainty: Decoupling Capability and Calibration in On-Policy Distillation , author=. arXiv preprint arXiv:2604.16830 , year=

    Pith/arXiv arXiv

  45. [45]

    arXiv preprint arXiv:2604.13016 , year=

    Rethinking On-Policy Distillation of Large Language Models: Phenomenology, Mechanism, and Recipe , author=. arXiv preprint arXiv:2604.13016 , year=

    Pith/arXiv arXiv

  46. [46]

    arXiv preprint arXiv:2601.18734 , year=

    Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models , author=. arXiv preprint arXiv:2601.18734 , year=

    Pith/arXiv arXiv

  47. [47]

    arXiv preprint arXiv:2601.20802 , year=

    Reinforcement Learning via Self-Distillation , author=. arXiv preprint arXiv:2601.20802 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2603.23871 , year=

    HDPO: Hybrid Distillation Policy Optimization via Privileged Self-Distillation , author=. arXiv preprint arXiv:2603.23871 , year=

    arXiv

  49. [49]

    DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence , author=

  50. [50]

    arXiv preprint arXiv:2601.02780 , year=

    Mimo-v2-flash technical report , author=. arXiv preprint arXiv:2601.02780 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2604.03128 , year=

    Self-Distilled RLVR , author=. arXiv preprint arXiv:2604.03128 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2604.14084 , year=

    TIP: Token Importance in On-Policy Distillation , author=. arXiv preprint arXiv:2604.14084 , year=

    Pith/arXiv arXiv

  53. [53]

    arXiv preprint arXiv:2505.21523 , year=

    More thinking, less seeing? assessing amplified hallucination in multimodal reasoning models , author=. arXiv preprint arXiv:2505.21523 , year=

    Pith/arXiv arXiv

  54. [54]

    arXiv preprint arXiv:2507.06448 , year=

    Perception-aware policy optimization for multimodal reasoning , author=. arXiv preprint arXiv:2507.06448 , year=

    Pith/arXiv arXiv

  55. [55]

    arXiv preprint arXiv:2604.08541 , year=

    Seeing but Not Thinking: Routing Distraction in Multimodal Mixture-of-Experts , author=. arXiv preprint arXiv:2604.08541 , year=

    Pith/arXiv arXiv

  56. [56]

    arXiv preprint arXiv:2601.21821 , year=

    MMFineReason: Closing the Multimodal Reasoning Gap via Open Data-Centric Methods , author=. arXiv preprint arXiv:2601.21821 , year=

    arXiv

  57. [57]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  58. [58]

    Proceedings of the Twentieth European Conference on Computer Systems , pages=

    Hybridflow: A flexible and efficient rlhf framework , author=. Proceedings of the Twentieth European Conference on Computer Systems , pages=

  59. [59]

    Proceedings of the 29th symposium on operating systems principles , pages=

    Efficient memory management for large language model serving with pagedattention , author=. Proceedings of the 29th symposium on operating systems principles , pages=

  60. [60]

    International Conference on Learning Representations , volume=

    Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts , author=. International Conference on Learning Representations , volume=

  61. [61]

    Advances in Neural Information Processing Systems , volume=

    Measuring multimodal mathematical reasoning with math-vision dataset , author=. Advances in Neural Information Processing Systems , volume=

  62. [62]

    International Conference on Learning Representations , volume=

    Dynamath: A dynamic visual benchmark for evaluating mathematical reasoning robustness of vision language models , author=. International Conference on Learning Representations , volume=

  63. [63]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    We-math: Does your large multimodal model achieve human-like mathematical reasoning? , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  64. [64]

    Advances in Neural Information Processing Systems , volume=

    Charxiv: Charting gaps in realistic chart understanding in multimodal llms , author=. Advances in Neural Information Processing Systems , volume=

  65. [65]

    Findings of the association for computational linguistics: ACL 2022 , pages=

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning , author=. Findings of the association for computational linguistics: ACL 2022 , pages=

    2022

  66. [66]

    arXiv preprint arXiv:2504.15279 , year=

    Visulogic: A benchmark for evaluating visual reasoning in multi-modal large language models , author=. arXiv preprint arXiv:2504.15279 , year=

    Pith/arXiv arXiv

  67. [67]

    arXiv preprint arXiv:2407.04973 , year=

    Logicvista: Multimodal llm logical reasoning benchmark in visual contexts , author=. arXiv preprint arXiv:2407.04973 , year=

    Pith/arXiv arXiv

  68. [68]

    Science China Information Sciences , volume=

    Ocrbench: on the hidden mystery of ocr in large multimodal models , author=. Science China Information Sciences , volume=. 2024 , publisher=

    2024

  69. [69]

    Advances in Neural Information Processing Systems , volume=

    Are we on the right way for evaluating large vision-language models? , author=. Advances in Neural Information Processing Systems , volume=

  70. [70]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Hallusionbench: an advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  71. [71]

    arXiv preprint arXiv:2605.11739 , year=

    Learning to Foresee: Unveiling the Unlocking Efficiency of On-Policy Distillation , author=. arXiv preprint arXiv:2605.11739 , year=

    Pith/arXiv arXiv

  72. [72]

    arXiv preprint arXiv:2510.00553 , year=

    On Predictability of Reinforcement Learning Dynamics for Large Language Models , author=. arXiv preprint arXiv:2510.00553 , year=

    arXiv

  73. [73]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    Pith/arXiv arXiv

  74. [74]

    arXiv preprint arXiv:2605.03677 , year=

    Uni-OPD: Unifying on-policy distillation with a dual-perspective recipe , author=. arXiv preprint arXiv:2605.03677 , year=

    Pith/arXiv arXiv