Pith. sign in

REVIEW 3 major objections 6 minor 39 references

OPOD: On-Policy Omni Distillation

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read OPOD claims that routing a student's own responses to the matching text, image, or audio teacher, while keeping only teacher-ahead token guidance, yields one omni-modal model that outperforms every specialist and pooled multimodal training.

desk verdict A solid, internally consistent multi-teacher distillation recipe for omni-modal models; the routing worry in the stress-test doesn't land, but missing error bars and code do. read the letter →

arxiv 2607.20918 v1 pith:RS7DXSPP submitted 2026-07-23 cs.AI

Tong Zhao , Yuyang Hu , Reed Li , Yu Lu , Haibo Shi , Yutao Zhu , Zhicheng Dou This is my paper · ORCID
classification cs.AI
keywords on-policydistillationomni-modalmodelmulti-teacherone-sidedtokenguidanceadaptivemodalitycontrolverificationrewardcross-modalbalanceprocess
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 tries to establish that a single omni-modal model can improve across text, image, and audio at once, rather than trading one ability for another, by distilling from three modality-specialist teachers instead of training on pooled data. The proposed method, OPOD, has the student generate its own responses, routes each response to the matching specialist teacher, and keeps teacher guidance only on tokens where the teacher is more confident than the student. It also gives each modality its own adjustable guidance budget and lets the routed teacher verify both the final answer and the reasoning behind it. Across twelve benchmarks and three backbone sizes, the student reaches the best average score at every scale, up to 70.8 on the largest model, exceeding the strongest comparator by 2.1 points; on that model it even beats a per-benchmark oracle that picks the best specialist for each task.

What carries the argument

The central object is the one-sided token margin $c_{i,t} = \mathrm{ReLU}(r^{\mathrm{opd}}_{i,t})$, where $r^{\mathrm{opd}}_{i,t} = \log \pi_T^{m_i}(y_{i,t} \mid x_i, y_{i,<t}) - \log \pi_\theta(y_{i,t} \mid x_i, y_{i,<t})$ is the teacher-student log-ratio on a token the student actually sampled. This ReLU keeps teacher guidance only where the routed teacher is ahead, so the constraint vanishes once the student matches or exceeds the teacher. Aggregated into a per-rollout constraint $\bar{c}_m$, it drives a dual controller that updates each modality's weight $\beta_m$ toward a separately calibrated budget $\epsilon_m$, preventing one teacher from dominating the shared backbone. The same teacher also acts as a trajectory verifier, producing a process reward that combines correctness-gated answer confidence with an ungated reasoning-gain term. These components are joined in one objective $L = L_{\mathrm{rl}} + L_{\mathrm{tc}} + \alpha L_{\mathrm{ref}}$, which updates the shared student while the teachers remain fixed.

What would settle it

Rerun OPOD on the omni-modal benchmark with the same training pipeline but three different routing rules for mixed prompts (text-majority, first-modality, oracle-assigned); if the 30B OmniBench score (reported 61.1) shifts by more than about one point across rules, the single-label routing premise is load-bearing and the omni-modal gain is not yet pinned down.

Watch

Extended reading notes

Core claim

The central claim, stated on the paper's own terms, is that coordinating modality-specific teachers through three complementary mechanisms makes one omni-modal student better than any specialist and better than pooled-data post-training. The student samples its own rollouts; each rollout is routed by its modality label to the matching text, image, or audio teacher, which was trained separately with the same policy-gradient recipe. The teacher's evaluation is converted into three coordinated signals: one-sided token guidance that keeps only the positive teacher-student log-probability margin, a per-modality adaptive controller that maintains a separate trust-region weight and constraint budget for each modality, and a verification reward in which the routed teacher scores answer confidence and reasoning gain over the whole trajectory. On the 30B backbone, the student reaches a 70.8 average, surpassing the strongest individual specialist by 2.1 points and a per-benchmark oracle by 0.5 points, while the three specialists are discarded after training, leaving one deployable omni-modal policy.

Load-bearing premise

Every prompt, including mixed text-image-audio ones, is assumed to carry a single known modality label that selects exactly one teacher, but the paper never states how mixed prompts get that label.

Editorial extensions

If this is right

  • The specialists can be discarded after training, so the deployed model is a single omni-modal policy with no ensemble latency or memory overhead.
  • The method improves all modality groups at once on the 30B backbone, unlike pooled GRPO (which drops the omni-modal score), native OPD (which drops vision), and ExOPD (which drops audio and omni-modal).
  • The gains persist across three backbone scales and two model families, with the margin over the strongest alternative growing from 1.7 points at 3B to 2.2 points at 30B.
  • A per-benchmark oracle that picks the best specialist still trails the single OPOD student by 0.5 points on the 30B model, so consolidation is doing more than selecting the strongest teacher per task.
  • Ablations show each of the three coordinated components (one-sided guidance, adaptive modality control, verification reward) contributes between 1.4 and 2.2 points, with per-modality control producing the largest degradation when removed.

Reading between the lines

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

  • Not stated in the paper: mixed text-image-audio prompts are assumed to have one clean modality label; a direct extension is to test whether a learned router or a majority-vote rule changes the omni-modal results.
  • The paper does not explore why the single student beats the per-benchmark oracle; a natural follow-up is to isolate which tokens each teacher contributes, since the one-sided margin records exactly where each teacher was ahead.
  • Because one-sided ReLU guidance only keeps teacher-ahead tokens, the same signal could double as a confidence-weighted training filter or an early-stopping rule; this is an editorial extension, not a paper claim.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes OPOD, a framework for distilling modality-specific teachers (text, image, and audio) into a single omni-modal student via on-policy distillation. Student rollouts are routed by the prompt's modality label to the corresponding GRPO-trained teacher, and the training objective combines three components: one-sided token guidance (ReLU of the teacher-student log-ratio), per-modality dual control with budgets estimated from a warm-up, and a verification reward that uses the routed teacher to score answer confidence and reasoning gain. The authors evaluate on Qwen3-Omni-30B-A3B and Qwen2.5-Omni-7B/3B across 12 benchmarks, reporting the best average at each scale (70.8, 51.7, and 46.2 points) and ablations attributing 1.4 to 2.2 points to each component.

Significance. If the reported results are reproducible, OPOD is a practical recipe for consolidating modality specialists into a single deployable omni-modal model, and the three-level design is well motivated by the observed cross-modal update conflicts. Strengths include the component ablations, the cross-scale transfer to dense 3B/7B backbones, the per-benchmark oracle comparison (OPOD 70.8 vs. oracle 70.3), and the controller dynamics analysis. The main caveats are the under-specified routing for omni-modal prompts and the absence of multiple seeds, given that the headline margins over the strongest comparator are only 1.7 to 2.1 points.

major comments (3)
  1. [Preliminaries / Table 1] The paper defines training prompts as (x_i, m_i) with m_i in {text, audio, image} and routes each rollout to the teacher pi_T^{m_i} (Eq. (1) and Figure 2), but the evaluation includes OmniBench, an omni-modal benchmark (Table 1). The paper never states whether OmniBench prompts are included in distillation training and, if so, how a mixed-modality prompt is assigned a single modality label. Since no teacher is omni-modal, Eq. (5) is undefined for such a prompt absent an explicit routing rule. If OmniBench is used only as held-out evaluation, this should be stated explicitly; as written, the reader cannot determine whether the reported OmniBench column reflects the method as specified or a hidden routing heuristic. This gap directly affects the headline 12-benchmark average.
  2. [Experimental Setup / Table 1] The reported margins over the strongest comparator are small (1.7, 1.8, and 2.1 points), and the paper provides no error bars, no multiple seeds, and no significance tests. Several per-benchmark differences are within 1 to 2 points (e.g., 30B AIME26: OPOD 63.3 vs. ExOPD 64.2; 7B MMLU-Pro: OPOD 52.5 vs. ExOPD 52.4). Without any estimate of run-to-run variance, the central claim that OPOD achieves the best average at every scale is not robustly established. At minimum, the revision should report results across at least two seeds (preferably three) with standard deviations for the main table and the ablations.
  3. [Adaptive Modality Control / Figure 4] The per-modality budgets epsilon_m are central to the adaptive control component that yields the largest ablation drop (2.2 points), but the estimation procedure is not fully specified. The text says the budget is "estimated once from the mean warm-up constraint for that modality," without stating the estimator (e.g., batch-wise mean vs. exponential moving average), the window size, or how absent modalities are handled. Figure 4(c) reports an image budget of 0.191 while the caption notes that the final warm-up batch contains no image sample; it is unclear how the running estimate is updated in this case. Please specify the exact warm-up computation and report sensitivity to the warm-up length and the floor epsilon_min.
minor comments (6)
  1. [Table 1] The rows contain run-together numbers (e.g., "57.741.576.8" for the text teacher and "50.064.241.0" for ExOPD); please format the columns with clear separators.
  2. [Preliminaries] The set of modalities is written {text, audio, image} in one place and {text, image, audio} in others; standardize the ordering throughout, including the Abstract.
  3. [Figure 4 caption] Please explain why a frozen image budget is reported even though the final warm-up batch has no image sample, and state how the running estimate decays or carries over for absent modalities.
  4. [Experimental Setup] The initial values of beta_m, the clip bounds beta_min and beta_max, and the dual step size eta are not given for the main experiments; Figure 4 reports beta_min=0.1 but the text only says 0 <= beta_min < beta_max. Please report these hyperparameters.
  5. [Equation (2)] The definition of c_i assumes a positive response length; clarify the handling of empty or degenerate rollouts, or state that such rollouts are filtered.
  6. [Baselines] The "Native OPD" baseline is described as standard on-policy distillation, but the large vision drop (ChartQA 73.4 vs. base 85.8 in Table 1) suggests the exact objective differs from the one-sided version; please specify the baseline loss (e.g., symmetric KL or a different margin) to ensure a fair comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: OPOD's benchmark results are measured externally after training, and the adaptive budgets and reward weights are training hyperparameters rather than inputs that compute the reported scores.

full rationale

The paper's central claim is an empirical comparison on twelve external benchmarks after post-training, not a derivation whose output is forced by its inputs. The training objective L = L_rl + L_tc + alpha L_ref is a standard on-policy loss with a teacher-constraint term and a reference KL regularizer; the reported accuracies in Table 1 are measured after training and are not presented as predictions computed from Eq. (5). The per-modality budgets epsilon_m are estimated from a 10-step warm-up of the student's own one-sided constraints, but this only sets the target for the dual controller; it does not by construction determine the benchmark numbers, and the paper does not claim that it does. One-sided guidance, modality control, and verification reward are design choices evaluated by ablations rather than definitions whose equivalence to the measured outcomes is asserted. The only substantive issue identified, the unspecified routing rule for genuinely mixed-modality prompts such as those in OmniBench, is a specification gap: the equations define routing only for m_i in {text, audio, image}, and the paper does not state whether OmniBench prompts appear in distillation training or how they would be labeled. This gap affects clarity and reproducibility but does not make any result logically equivalent to its inputs. The self-citations present, such as the OmniGAIA reference in Related Work, are background context and are not load-bearing for the central claim. Therefore no derivation-level circularity is exhibited.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central empirical claim rests on standard RL/KD background (GRPO, reverse-KL distillation), on the assumption that specialist teachers trained on separate modality data are complementary and transferable to one shared backbone, and on the assumption that every prompt has a single known modality label. The only data-fitted numbers are the per-modality epsilon budgets (0.020, 0.191, 0.254) estimated from the student's own warm-up constraints; several other hyperparameters are unreported.

free parameters (5)
  • Modality constraint budgets epsilon_text, epsilon_image, epsilon_audio = 0.020, 0.191, 0.254
    Estimated from the mean one-sided constraint during a 10-step warm-up and frozen (Section 'Adaptive Modality Control', Figure 4a). They set each modality's target constraint and directly control the dual weight update in Eq. 3.
  • Initial dual weights beta_m and clip bounds = beta_init about 0.6, beta_min 0.1, beta_max unreported
    All three per-modality teacher weights start near 0.6 and are clipped to [0.1, beta_max]; these hand-chosen values govern how much teacher pressure is applied (Figure 4c).
  • Reward weights wA, wB and clip C = 0.2, 0.1, 2.0
    Hand-set weights for answer confidence and reasoning gain in Eq. 4; they shape the verification reward.
  • Reference KL weight alpha = unreported
    The objective L = L_rl + L_tc + alpha L_ref uses a reference KL regularizer; its value is not given, so the objective is incompletely specified.
  • Dual step size eta and floor eps_min = eta unreported; eps_min=0.02
    eta controls the adaptivity of beta_m in Eq. 3; eps_min lower-bounds the estimated budgets.
assumptions (5)
  • domain assumption Specialist teachers trained with GRPO on separate modality data show complementary strengths that can be consolidated into one student that beats each specialist on most benchmarks.
    The motivating diagnostic in Figure 1(a) shows complementary strengths, but the assumption that a single shared backbone can retain all of them is the empirical basis of the method.
  • domain assumption The one-sided ReLU margin keeps only teacher-student log-ratio where the teacher is ahead, and this is a sufficient token-level guidance signal.
    Eq. 2 drops negative margins, assuming the student should not be pulled back once it exceeds the teacher; this is a modeling choice rather than a proven optimality result.
  • domain assumption The reasoning gain B_i, the teacher's log-probability of the gold answer given the student's reasoning, is a valid reward for reasoning quality even when the final answer is wrong.
    Section 'Verification Reward' makes this ungated by correctness; it assumes teacher likelihood shifts are informative without human step annotations.
  • domain assumption Each prompt has a known modality label m_i that uniquely selects one teacher, including omni-modal inputs.
    Preliminaries defines routing by 'the known modality label m_i'; the paper never specifies how mixed-modality OmniBench prompts are labeled, yet reports omni-modal results.
  • standard math GRPO and the reverse-KL interpretation of on-policy distillation are valid optimization objectives.
    Taken as background from Shao et al. 2024 and Agarwal et al. 2024; the paper does not re-derive them.

how reviews work

0 comments
Cite this review

Pith. "Pith review of OPOD: On-Policy Omni Distillation." pith.science (2026). https://pith.science/paper/RS7DXSPP

@misc{pith2026260720918,
  author       = {Pith},
  title        = {Pith review of: OPOD: On-Policy Omni Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RS7DXSPP}},
  note         = {Machine review of arXiv:2607.20918}
}
read the original abstract

Omni-modal models can handle text, images, and audio in one system, but improving all of these abilities together remains difficult. Training a single model on pooled multimodal data often fails to match models specialized for individual modalities. On-policy distillation (OPD) offers a way to combine such specialists: the student generates a response, and a teacher evaluates that same response, so the student learns directly from behaviors it actually produces. Yet using several teachers can introduce competing guidance and improve one modality at the expense of another. We present On-Policy Omni Distillation (OPOD), which routes each student response to the matching text, image, or audio teacher. OPOD keeps teacher guidance only on tokens where the teacher assigns a higher probability than the student, adjusts the influence of each modality teacher independently during training, and asks the routed teacher to assess both the final answer and whether the reasoning supports the correct answer. Across twelve benchmarks and three backbone sizes, OPOD achieves the best average score at every scale, reaching 70.8, 51.7, and 46.2 and exceeding the strongest comparator by 2.1, 1.8, and 1.7 points. On the 30B model, it outperforms both the base model and a counterpart post-trained jointly on pooled multimodal data on all twelve benchmarks, and ranks first or second on eleven even when the individual specialists are included. The specialists are discarded after training, leaving one deployable omni-modal model. These results show that coordinating modality-specific teachers is an effective way to improve a shared model while maintaining cross-modal balance.

Figures

Figures reproduced from arXiv: 2607.20918 by the authors.

Figure 1
Figure 1. Motivating observations for OPOD. (a) With GRPO held fixed, text, image, and audio specialists exhibit complementary strengths, whereas a model trained on pooled data does not consistently match them. These specialists subsequently serve as teachers. (b) The same teachers show frequent pairwise conflicts in their parameter-displacement directions relative to the common base model. (c) OPOD produces broader benchmark… view at source ↗
Figure 2
Figure 2. Overview of OPOD. Student rollouts are routed by input modality to the corresponding teacher trained offline with GRPO. OPOD combines one-sided token guidance, modality-specific constraint control, and teacher-based verification through answer confidence and reasoning gain. The resulting policy loss, teacher constraint, and reference regularizer jointly update the shared student. On-Policy Omni Distillation Prelimin… view at source ↗
Figure 3
Figure 3. Improvement in the 12-benchmark average over the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Dynamics of adaptive modality control in the Qwen3-Omni-30B-A3B [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 15 canonical work pages

  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]

    On-policy distillation of language models: Learning from self-generated mistakes

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos Garea, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. https://openreview....

    work page 2024

  3. [3]

    Qwen-vl: A frontier large vision-language model with versatile abilities

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. CoRR, abs/2308.12966, 2023. doi:10.48550/ARXIV.2308.12966. https://doi.org/10.48550/arXiv.2308.12966

    arXiv doi:10.48550/arxiv.2308.12966 2023

  4. [4]

    Qwen2-audio technical report

    Yunfei Chu, Jin Xu, Qian Yang, Haojie Wei, Xipin Wei, Zhifang Guo, Yichong Leng, Yuanjun Lv, Jinzheng He, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen2-audio technical report. CoRR, abs/2407.10759, 2024. doi:10.48550/ARXIV.2407.10759. https://doi.org/10.48550/arXiv.2407.10759

    arXiv doi:10.48550/arxiv.2407.10759 2024

  5. [5]

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven C. H. Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning. In Alice Oh, Tristan Naumann, Amir Globerson, Kate Saenko, Moritz Hardt, and Sergey Levine, editors, Advances in Neural Information Proc...

    work page 2023

  6. [6]

    Minillm: Knowledge distillation of large language models

    Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. https://openreview.net/forum?id=5h0qf7IBZZ

    work page 2024

  7. [7]

    Hinton, Oriol Vinyals, and Jeffrey Dean

    Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015. http://arxiv.org/abs/1503.02531

    arXiv 2015

  8. [8]

    Uni-opd: Unifying on-policy distillation with a dual-perspective recipe

    Wenjin Hou, Shangpin Peng, Weinong Wang, Zheng Ruan, Yue Zhang, Zhenglin Zhou, Mingqi Gao, Yifei Chen, Kaiqi Wang, Hongming Yang, Chengquan Zhang, Zhuotao Tian, Han Hu, Yi Yang, Fei Wu, and Hehe Fan. Uni-opd: Unifying on-policy distillation with a dual-perspective recipe. CoRR, abs/2605.03677, 2026. doi:10.48550/ARXIV.2605.03677. https://doi.org/10.48550/...

    arXiv doi:10.48550/arxiv.2605.03677 2026

Show all 39 references
  1. [9]

    Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In Jian Su, Xavier Carreras, and Kevin Duh, editors, Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, EMNLP 2016, Austin, Texas, USA, November 1-4, 2016 , pages 1317--...

    2016 doi

  2. [10]

    Video-opd: Efficient post-training of multimodal large language models for temporal video grounding via on-policy distillation

    Jiaze Li, Hao Yin, Haoran Xu, Boshen Xu, Wenhui Tan, Zewen He, Jianzhong Ju, Zhenbo Luo, and Jian Luan. Video-opd: Efficient post-training of multimodal large language models for temporal video grounding via on-policy distillation. CoRR, abs/2602.02994, 2026 a . doi:10.48550/A...

    2026 doi arXiv

  3. [11]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven C. H. Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In Andreas Krause, Emma Brunskill, Kyunghyun Cho, Barbara Engelhardt, Sivan Sabato, and Jonathan Scarlett, edito...

    2023

  4. [12]

    Omnigaia: Towards native omni-modal AI agents

    Xiaoxi Li, Wenxiang Jiao, Jiarui Jin, Shijian Wang, Guanting Dong, Jiajie Jin, Hao Wang, Yinuo Wang, Ji - Rong Wen, Yuan Lu, and Zhicheng Dou. Omnigaia: Towards native omni-modal AI agents. CoRR, abs/2602.22897, 2026 b . doi:10.48550/ARXIV.2602.22897. https://doi.org/10.48550/...

    2026 doi arXiv

  5. [13]

    Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe

    Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan - ang Gao, Wenkai Yang, Zhiyuan Liu, and Ning Ding. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe. CoRR, abs/2604.13016, 2026 c . doi...

    2026 doi arXiv

  6. [14]

    Omnibench: Towards the future of universal omni-language models

    Yizhi Li, Ge Zhang, Yinghao Ma, Ruibin Yuan, Hangyu Guo, Yiming Liang, Jiaheng Liu, Noah Wang, Jian Yang, Siwei Wu, et al. Omnibench: Towards the future of universal omni-language models. Advances in Neural Information Processing Systems, 38, 2026 d

    2026

  7. [15]

    Let's verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let's verify step by step. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-1...

    2024

  8. [16]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. CoRR, abs/2304.08485, 2023. doi:10.48550/ARXIV.2304.08485. https://doi.org/10.48550/arXiv.2304.08485

    2023 doi arXiv

  9. [17]

    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 International Conference on Learning Representat...

    2024

  10. [18]

    MOPD: multi-teacher on-policy distillation for capability integration in LLM post-training

    Wenhan Ma, Jianyu Wei, Liang Zhao, Hailin Zhang, Bangjun Xiao, Lei Li, Qibin Yang, Bofei Gao, Yudong Wang, Rang Li, Jinhao Dong, Zhifang Sui, and Fuli Luo. MOPD: multi-teacher on-policy distillation for capability integration in LLM post-training. CoRR, abs/2606.30406, 2026. d...

    2026 doi arXiv

  11. [19]

    Chartqa: A benchmark for question answering about charts with visual and logical reasoning

    Ahmed Masry, Jia Qing Tan, Shafiq Joty, Enamul Hoque, et al. Chartqa: 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, 2022

    2022

  12. [20]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level google-proof q & a benchmark. CoRR, abs/2311.12022, 2023. doi:10.48550/ARXIV.2311.12022. https://doi.org/10.48550/ar...

    2023 doi arXiv

  13. [21]

    Mmau: A massive multi-task audio understanding and reasoning benchmark

    Sakshi Sakshi, Utkarsh Tyagi, Sonal Kumar, Ashish Seth, Ramaneswaran Selvakumar, Oriol Nieto, Ramani Duraiswami, Sreyan Ghosh, and Dinesh Manocha. Mmau: A massive multi-task audio understanding and reasoning benchmark. In International Conference on Learning Representations, v...

    2025

  14. [22]

    Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter

    Victor Sanh, Lysandre Debut, Julien Chaumond, and Thomas Wolf. Distilbert, a distilled version of BERT: smaller, faster, cheaper and lighter. CoRR, abs/1910.01108, 2019. http://arxiv.org/abs/1910.01108

    1910 arXiv

  15. [23]

    A-okvqa: A benchmark for visual question answering using world knowledge

    Dustin Schwenk, Apoorv Khandelwal, Christopher Clark, Kenneth Marino, and Roozbeh Mottaghi. A-okvqa: A benchmark for visual question answering using world knowledge. In European conference on computer vision, pages 146--162. Springer, 2022

    2022

  16. [24]

    Rewarding progress: Scaling automated process verifiers for LLM reasoning

    Amrith Setlur, Chirag Nagpal, Adam Fisch, Xinyang Geng, Jacob Eisenstein, Rishabh Agarwal, Alekh Agarwal, Jonathan Berant, and Aviral Kumar. Rewarding progress: Scaling automated process verifiers for LLM reasoning. In The Thirteenth International Conference on Learning Repres...

    2025

  17. [25]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, Y. K. Li, Y. Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. CoRR, abs/2402.03300, 2024. doi:10.48550/ARXIV.2402.03300. https://doi.org/10.48550...

    2024 doi arXiv

  18. [26]

    Knowledge fusion of large language models

    Fanqi Wan, Xinting Huang, Deng Cai, Xiaojun Quan, Wei Bi, and Shuming Shi. Knowledge fusion of large language models. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024 . OpenReview.net, 2024. https://openreview.net/...

    2024

  19. [27]

    PRISM: pre-alignment via black-box on-policy distillation for multimodal reinforcement learning

    Sudong Wang, Weiquan Huang, Xiaomin Yu, Zuhao Yang, Hehai Lin, Keming Wu, Chaojun Xiao, Chen Chen, Wenxuan Wang, Beier Zhu, Yunjian Zhang, and Chengwei Qin. PRISM: pre-alignment via black-box on-policy distillation for multimodal reinforcement learning. CoRR, abs/2604.28123, 2...

    2026 doi arXiv

  20. [28]

    Mmlu-pro: A more robust and challenging multi-task language understanding benchmark

    Yubo Wang, Xueguang Ma, Ge Zhang, Yuansheng Ni, Abhranil Chandra, Shiguang Guo, Weiming Ren, Aaran Arulraj, Xuan He, Ziyan Jiang, et al. Mmlu-pro: A more robust and challenging multi-task language understanding benchmark. Advances in Neural Information Processing Systems, 37: ...

    2024

  21. [29]

    Qwen2.5-omni technical report

    Jin Xu, Zhifang Guo, Jinzheng He, Hangrui Hu, Ting He, Shuai Bai, Keqin Chen, Jialin Wang, Yang Fan, Kai Dang, Bin Zhang, Xiong Wang, Yunfei Chu, and Junyang Lin. Qwen2.5-omni technical report. CoRR, abs/2503.20215, 2025 a . doi:10.48550/ARXIV.2503.20215. https://doi.org/10.48...

    2025 doi arXiv

  22. [30]

    Qwen3-omni technical report

    Jin Xu, Zhifang Guo, Hangrui Hu, Yunfei Chu, Xiong Wang, Jinzheng He, Yuxuan Wang, Xian Shi, Ting He, Xinfa Zhu, et al. Qwen3-omni technical report. arXiv preprint arXiv:2509.17765, 2025 b

    2025 arXiv

  23. [31]

    Avqa: A dataset for audio-visual question answering on videos

    Pinci Yang, Xin Wang, Xuguang Duan, Hong Chen, Runze Hou, Cong Jin, and Wenwu Zhu. Avqa: A dataset for audio-visual question answering on videos. In Proceedings of the 30th ACM international conference on multimedia, pages 3480--3491, 2022

    2022

  24. [32]

    Learning beyond teacher: Generalized on-policy distillation with reward extrapolation

    Wenkai Yang, Weijie Liu, Ruobing Xie, Kai Yang, Saiyong Yang, and Yankai Lin. Learning beyond teacher: Generalized on-policy distillation with reward extrapolation. CoRR, abs/2602.12125, 2026. doi:10.48550/ARXIV.2602.12125. https://doi.org/10.48550/arXiv.2602.12125

    2026 doi arXiv

  25. [33]

    Cohen, Ruslan Salakhutdinov, and Christopher D

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W. Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. CoRR, abs/1809.09600, 2018. http://arxiv.org/abs/1809.09600

    2018 arXiv

  26. [34]

    Black-box on-policy distillation of large language models

    Tianzhu Ye, Li Dong, Zewen Chi, Xun Wu, Shaohan Huang, and Furu Wei. Black-box on-policy distillation of large language models. CoRR, abs/2511.10643, 2025. doi:10.48550/ARXIV.2511.10643. https://doi.org/10.48550/arXiv.2511.10643

    2025 doi

  27. [35]

    H-opd: Confidence aware heterogeneous multi-teacher multimodal on-policy distillation, 2026

    Qixiang Yin, Huanjin Yao, Yuchen Cai, Jianghao Chen, Ziyi Wang, Min Yang, Fei Su, and Zhicheng Zhao. H-opd: Confidence aware heterogeneous multi-teacher multimodal on-policy distillation, 2026. https://arxiv.org/abs/2607.02592

    2026 arXiv

  28. [36]

    Learning from multiple teacher networks

    Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada, August 13 - 17, 2017 , pages 1285--1294. ACM , 2017. doi:10.1145/3...

    2017

  29. [37]

    Vision-opd: Learning to see fine details for multimodal llms via on-policy self-distillation, 2026

    Qianhao Yuan, Jie Lou, Xing Yu, Hongyu Lin, Le Sun, Xianpei Han, and Yaojie Lu. Vision-opd: Learning to see fine details for multimodal llms via on-policy self-distillation, 2026. https://arxiv.org/abs/2605.18740

    2026 arXiv

  30. [38]

    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, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF conference on c...

    2024

  31. [39]

    MM-RLHF: the next step forward in multimodal LLM alignment

    Yifan Zhang, Tao Yu, Haochen Tian, Chaoyou Fu, Peiyan Li, Jianshu Zeng, Wulin Xie, Yang Shi, Huanyu Zhang, Junkang Wu, Xue Wang, Yibo Hu, Bin Wen, Tingting Gao, Zhang Zhang, Fan Yang, Di Zhang, Liang Wang, and Rong Jin. MM-RLHF: the next step forward in multimodal LLM alignmen...

    2025

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.