Pith. sign in

REVIEW 3 major objections 5 minor 98 references

Teacher corrections in multimodal distillation mix visual evidence with language priors; VAD keeps only the part the evidence can explain and rebuilds the learning target around the student.

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

2026-07-31 02:54 UTC pith:IGRTTSZZ

load-bearing objection Solid method paper: reconstruct the OPD target from a signed present/removed teacher contrast, with real matched-budget gains; the crop-vs-blur proxy is the caveat, not a collapse of the argument. the 3 major comments →

arxiv 2607.28590 v1 pith:IGRTTSZZ submitted 2026-07-30 cs.CV cs.CL

VAD: Attributing Visual Evidence for Target Reconstruction in Multimodal On-Policy Distillation

Kangning Zhang , Yixing Li , Shuai Shao , Qingyao Li , Zhengxi Lu , Zhiyuan Yao , Jianghao Lin , Wenxiang Jiao
show 4 more authors
Yuan Lu Weiwen Liu Weinan Zhang Yong Yu
This is my paper
classification cs.CV cs.CL
keywords multimodal on-policy distillationvisual attributioncounterfactual target reconstructionprivileged teacherfine-grained visual perceptionsupport and refutationknowledge distillation
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.

When a multimodal student is trained by matching a privileged teacher that sees a clearer crop of the image, the teacher’s next-token fixes are mixed: some come from the visual evidence, some from language habits and teacher quirks. This paper argues that the useful move is not only to decide where or how hard to distill, but to estimate which part of each correction the controlled visual change actually supports. Visual Attribution Distillation (VAD) does that by querying the same fixed teacher twice on each student-generated prefix—once with the evidence present and once with it removed—then treating the shift in centered log-probabilities as a signed proxy for the visual evidence direction. It projects the original teacher correction onto that proxy, rebuilds a student-anchored target from the aligned support and refutation, and uses the full privileged teacher only as a weak regularizer. On six fine-grained visual benchmarks at 4B and 9B scales, this reconstructed target beats direct privileged-view matching and visual-advantage weighting, and token analyses show the aligned component is enriched in visual, task-relevant corrections—especially when evidence should refute a wrong answer.

Core claim

Counterfactual target reconstruction outperforms source-mixed privileged supervision in multimodal on-policy distillation: projecting the teacher-to-student correction onto a same-teacher evidence-present versus evidence-removed centered log-probability proxy, then supervising with the student-anchored reconstructed target, yields stronger fine-grained visual accuracy than direct privileged-view distillation or visual-advantage weighting at both 4B and 9B scales.

What carries the argument

Visual Attribution Distillation (VAD): a signed proxy ut from the change in the fixed teacher’s centered log-probabilities when relevant evidence is present versus removed; a one-sided projection of the full teacher correction onto ut that isolates an intervention-aligned component; and budgeted support/refutation reconstruction of a student-anchored target from that component.

Load-bearing premise

A single marked-region crop pair—clear evidence versus a heavily downsampled version of the same crop—under unchanged text is enough to mark which teacher corrections the visual evidence truly warrants, rather than crop or degradation artifacts.

What would settle it

If, under matched data and budgets, reconstructing the target from the proxy-aligned component no longer beats direct matching of the evidence-present teacher and visual-advantage weighting on the six fine-grained benchmarks, or if token analyses no longer show enrichment of visual and correct-answer corrections in the aligned component, the central claim fails.

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

If this is right

  • Privileged multimodal OPD should reconstruct what to distill from a controlled evidence intervention, not only reweight full teacher matching.
  • Signed support and refutation matter: evidence that lowers a mistaken token can be as useful as evidence that raises the right one.
  • A weak full-teacher regularizer can stabilize language and formatting while the primary signal remains the visually attributed target.
  • Student-anchored targets built this way can improve fine-grained perception without sacrificing held-out generalization relative to the base model.

Where Pith is reading between the lines

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

  • Any privileged teacher channel—region crops, tools, or future text—may need the same present-versus-absent attribution step before its corrections are trusted as targets.
  • Richer multi-view or learned directional bases could shrink the residual that still mixes visual and nonvisual teacher effects.
  • The same counterfactual projection idea could test whether other distillation gains come from true evidence transfer or from teacher-specific language style.

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 / 5 minor

Summary. The paper proposes Visual Attribution Distillation (VAD) for multimodal on-policy distillation. It argues that privileged-view teacher corrections are source-mixed and that the key problem is estimating which part of a correction is supported by visual evidence. VAD queries a fixed teacher on evidence-present and evidence-removed views of the same student prefix, forms a signed centered log-probability proxy ut, one-sided-projects the privileged correction rt onto ut, budgets support versus refutation, and reconstructs a student-anchored target q_VAD that supplies primary JSD supervision, with a weak privileged-teacher regularizer. On six fine-grained benchmarks with matched Qwen3.5 4B/9B data and budgets, VAD outperforms Vision-OPD, VA-OPD, V-Zero, Decomposed OPD, and GRPO; ablations, semantic analyses of r_vis versus r_res, offline answer-token effects, held-out checks, and hyperparameter sweeps support the target-reconstruction story.

Significance. If the result holds under the intended interpretation, this is a clear and useful contribution to multimodal OPD: it reframes supervision from where/how strongly to distill toward reconstructing what to distill, with signed support and refutation. The empirical package is comparatively strong for the area—scale-matched post-training, six primary benchmarks, held-out generalization, target ablations (Table 2), offline correct/wrong effects (Fig. 4), semantic separation (Fig. 3), efficiency notes, and released code/models. Even under a weaker reading (a particular crop-versus-degradation contrast yields a better OPD target), the method and evidence would still matter for fine-grained MLLM post-training. The honesty that ut is a proxy, not an oracle, and the explicit residual caveats are strengths.

major comments (3)
  1. [§3.1–3.2, §4.1, Eqs. (3)–(7)] §3.1–3.2 and §4.1: the central attribution claim rests on ut = ϕ(p+_T) − ϕ(p−_T) from a pre-marked 2× crop (x+) versus the same crop after 0.1× bilinear downsampling and nearest-neighbor upsampling (x−). That pair confounds “evidence present/removed” with resolution, blur, and crop framing. One-sided projection (Eq. 4) and support/refutation budgeting (Eqs. 6–7) will inherit whatever that contrast encodes. Fig. 3 enrichment and Fig. 4 offline shifts are consistent with useful signal but do not separate visual evidence from degradation/crop artifacts. Please add at least one controlled intervention variant (e.g., mask/inpaint, alternate blur without crop change, or full-image vs masked full-image under matched framing) or materially soften claims that VAD “attributes visual evidence” versus “reconstructs targets from this specific contrastive proxy.”
  2. [Table 1, Table 2, Fig. 5] Table 1 and Table 2 report single-run accuracies without multi-seed means/stds or uncertainty on Avg6, while several margins are 0.2–3 points. Given free parameters (λ, τ+, ζ, c, top-K, degradation) and synthetic marked-region training, the headline ranking versus Vision-OPD/Decomposed OPD needs either repeated seeds for the main 4B/9B comparisons or a clear statement that results are single-run with sensitivity only via the λ/τ+ grid in Fig. 5. Without this, the load-bearing claim that target reconstruction—not run noise or tuning—drives the gains is under-supported.
  3. [§4.3, Fig. 3, Appendix A, §5] §4.3 / Fig. 3 and Limitations §5: the paper correctly says r_res is not purely linguistic and r_vis may retain nonvisual teacher effects, yet the narrative still treats proxy alignment as estimating the “visually attributable part” of the correction. The diagnostic in Appendix A also uses the same projection machinery used to define r_vis. Please keep motivation and evaluation distinct: either provide an external grounding check (e.g., human/attribute labels on high-|ut| tokens, or intervention swaps that should flip support/refutation) or revise abstract/intro/conclusion language so the empirical claim is “proxy-aligned reconstruction improves OPD” rather than identifiable visual attribution.
minor comments (5)
  1. [Figure 1] Figure 1 panels are dense; the 23.2%/22.8% diagnostic numbers are important but hard to parse next to the VA-OPD refutation cartoon. Consider a cleaner layout or moving numeric diagnostics fully to Appendix A with a short callout.
  2. [§3.1–3.2] Eq. (2)–(3): state explicitly whether Vt is always student top-K only or the union with teacher mass, and how the tail bucket enters centered log-probabilities; Appendix D helps but the main text should be self-contained.
  3. [Appendix C, Table 5] Table 5 efficiency comparison omits Decomposed OPD; either add it or note why timing logs were unavailable so the cost narrative is not selectively incomplete.
  4. [Throughout] Minor notation/consistency: r_VAD versus rvis, q_VAD_T,t versus qone, and occasional PDF artifacts (e.g., “/envel⌢pe”, “Te a c h e r”) should be cleaned for camera-ready.
  5. [§2.3] Related work §2.3: the contrast with Decomposed OPD is useful; a short explicit sentence on when multimodal-vs-text-only information gain and same-teacher evidence-present/removed projection would disagree would help readers.

Circularity Check

0 steps flagged

No significant circularity: VAD is an empirical training method whose accuracy claims are measured on held-out benchmarks, not forced by construction from its own definitions.

full rationale

The paper defines an intervention-derived proxy ut (Eq. 3), projects the privileged correction rt onto it (Eq. 4), reconstructs a student-anchored target (Eqs. 5–7), and trains with JSD plus a weak regularizer (Eqs. 8–10). Downstream claims are comparative accuracies on six external fine-grained benchmarks and held-out suites (Tables 1–4), plus semantic and offline target analyses (Figs. 3–4). None of these outcomes is algebraically equivalent to the inputs: the projection and target reconstruction specify a supervision signal; whether that signal improves Avg6 over Vision-OPD, VA-OPD, or Decomposed OPD is an empirical fact under matched data and budgets, not a fit renamed as prediction. Appendix A’s diagnostic A(m) reuses the same one-sided projection to motivate source-mixed corrections, which is standard problem framing rather than a load-bearing self-definition of the accuracy result. There is no uniqueness theorem, no fitted parameter presented as a first-principles forecast, and no self-citation chain that forces the central claim. Limitations explicitly treat ut as a proxy, not an oracle. Score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

As an empirical distillation paper, load-bearing structure is methodological assumptions plus hand-chosen training knobs, not physical axioms. The central claim rests on the intervention proxy, student-anchored projection geometry, weak regularizer, and the Vision-OPD synthetic crop protocol rather than on invented particles or fitted natural constants.

free parameters (6)
  • weak regularizer weight λ = 0.1
    Balances L_vis against privileged-teacher JSD; chosen as 0.1 after sweep in Fig. 5; full objective performance depends on it.
  • positive-branch cap τ+ = 0.8 (4B), 0.7 (9B)
    Caps support share of visual budget; set differently per scale (0.8 for 4B, 0.7 for 9B) via validation-style sweep.
  • projection stabilizer ζ = 1e-3
    Denominator ridge in one-sided projection β_t; fixed at 1e-3.
  • coordinate clip c = 20
    Bounds added logit shift before softmax; fixed at 20.
  • student top-K support size = 100 (+ tail)
    Shared candidate set V_t for centered log-probs and projection; training uses top-100 plus tail bucket.
  • evidence degradation factor = 0.1×
    x− built by 0.1× bilinear downsample then nearest upsample of the crop; defines the counterfactual contrast strength.
axioms (6)
  • ad hoc to paper Centered log-probability difference between evidence-present and evidence-removed teacher views is a useful signed proxy for the visual evidence direction at a prefix.
    Core construction in §3.1–3.2; authors explicitly say it is not an oracle grounding label.
  • ad hoc to paper One-sided projection of the privileged correction onto that proxy retains the visually warranted component and discards proxy-unexplained residual from primary supervision.
    Eq. (4)–(7); residual is acknowledged as still source-mixed, not purely linguistic.
  • domain assumption On-policy student prefixes plus a frozen same-initialization teacher are an appropriate distillation setting for transferring fine-grained perception.
    Inherited from OPD / Vision-OPD literature (§2.1, §4.1).
  • domain assumption Marked-region 2× crops and matched degraded crops in the Vision-OPD synthetic set adequately represent task-relevant visual evidence interventions.
    Training data protocol in §4.1; no training overlap with primary benchmarks claimed.
  • domain assumption Token-level Jensen–Shannon divergence on reconstructed targets is a suitable primary objective for sequence distillation.
    §3.4; Table 4 compares JSD vs forward/reverse KL.
  • domain assumption Official Vision-OPD inference pipeline with GPT-OSS-120B as judge faithfully ranks model accuracy on the six benchmarks.
    Evaluation protocol §4.1; shared across controlled rows.
invented entities (2)
  • visual evidence direction proxy u_t independent evidence
    purpose: Signed vector estimating how revealing evidence supports or refutes candidate tokens via teacher present/removed contrast.
    Defined in Eq. (3) from centered log-probs; operational construct, not a claimed physical latent, but it is the paper’s key postulated supervision object.
  • proxy-aligned visual correction r_vis / r_VAD and reconstructed target q_VAD independent evidence
    purpose: Primary distillation target replacing source-mixed privileged teacher distribution.
    Built by projection and budgeted support/refutation (§3.2–3.3); validated only via downstream training and token analyses in this paper’s setting.

pith-pipeline@v1.2.0-daily-grok45 · 26477 in / 4073 out tokens · 79684 ms · 2026-07-31T02:54:43.534743+00:00 · methodology

0 comments
read the original abstract

Multimodal on-policy distillation (OPD) transfers fine-grained visual knowledge by supervising student-generated trajectories with a privileged-view teacher. Yet its next-token corrections are source-mixed, combining visual signals with linguistic priors and teacher-specific effects. The key challenge is to estimate which corrections are supported by visual evidence, not merely where or how strongly to distill. We introduce Visual Attribution Distillation (VAD), a counterfactual target-reconstruction algorithm that estimates the visually attributable part of a teacher correction. At each student-generated prefix, VAD evaluates the same fixed teacher with the relevant evidence present and removed. The corresponding change in centered log-probabilities defines ut, a signed proxy for the visual evidence direction that estimates how revealing the evidence supports or refutes candidate tokens. VAD projects the original correction onto this proxy to obtain an intervention-aligned component and a proxy-unexplained residual, then reconstructs a student-anchored target from the former. During training, this reconstructed target supplies the primary supervision signal, while the privileged teacher contributes a weak regularizer. Across six fine-grained visual benchmarks at 4B and 9B scales, VAD outperforms direct privileged-view distillation and visual-advantage weighting. Token- level and controlled-target analyses show that the proxy-aligned component is enriched in task-relevant visual corrections and yields stronger target shifts, especially when evidence refutes a mistaken answer. These results support counterfactual target reconstruction as an effective alternative to source-mixed supervision.

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

98 extracted references · 35 linked inside Pith

  1. [1]

    2015 , eprint=

    Distilling the Knowledge in a Neural Network , author=. 2015 , eprint=

    2015

  2. [3]

    2024 , eprint=

    On-Policy Distillation of Language Models: Learning from Self-Generated Mistakes , author=. 2024 , eprint=

    2024

  3. [4]

    2023 , eprint=

    Visual Instruction Tuning , author=. 2023 , eprint=

    2023

  4. [5]

    2023 , eprint=

    InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning , author=. 2023 , eprint=

    2023

  5. [6]

    2023 , eprint=

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond , author=. 2023 , eprint=

    2023

  6. [7]

    2024 , eprint=

    Qwen2-VL: Enhancing Vision-Language Model's Perception of the World at Any Resolution , author=. 2024 , eprint=

    2024

  7. [8]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

    2025

  8. [9]

    2023 , eprint=

    V*: Guided Visual Search as a Core Mechanism in Multimodal LLMs , author=. 2023 , eprint=

    2023

  9. [10]

    2025 , eprint=

    MME: A Comprehensive Evaluation Benchmark for Multimodal Large Language Models , author=. 2025 , eprint=

    2025

  10. [11]

    2024 , eprint=

    Divide, Conquer and Combine: A Training-Free Framework for High-Resolution Image Perception in Multimodal Large Language Models , author=. 2024 , eprint=

    2024

  11. [12]

    2025 , eprint=

    MME-RealWorld: Could Your Multimodal LLM Challenge High-Resolution Real-World Scenarios that are Difficult for Humans? , author=. 2025 , eprint=

    2025

  12. [13]

    2026 , eprint=

    Zooming without Zooming: Region-to-Image Distillation for Fine-Grained Multimodal Perception , author=. 2026 , eprint=

    2026

  13. [14]

    2026 , eprint=

    Vision-OPD: Learning to See Fine Details for Multimodal LLMs via On-Policy Self-Distillation , author=. 2026 , eprint=

    2026

  14. [15]

    2026 , eprint=

    Visual-Advantage On-Policy Distillation for Vision-Language Models , author=. 2026 , eprint=

    2026

  15. [16]

    2026 , eprint=

    ViCuR: Visual Cues as Recoverable Privilege for Multimodal On-Policy Distillation , author=. 2026 , eprint=

    2026

  16. [17]

    2026 , eprint=

    DOPD: Dual On-policy Distillation , author=. 2026 , eprint=

    2026

  17. [18]

    2026 , eprint=

    Beyond Absolute Imitation: Anchored Residual Guidance for Privileged On-Policy Distillation , author=. 2026 , eprint=

    2026

  18. [19]

    2026 , eprint=

    Decomposed On-Policy Distillation for Vision-Language Reasoning: Steering Gradients for Visual Grounding , author=. 2026 , eprint=

    2026

  19. [20]

    2026 , eprint=

    V-Zero: Answer-Label-Free On-Policy Distillation with Contrastive Evidence Gating for Fine-Grained Visual Reasoning , author=. 2026 , eprint=

    2026

  20. [21]

    2024 , eprint=

    Eyes Wide Shut? Exploring the Visual Shortcomings of Multimodal LLMs , author=. 2024 , eprint=

    2024

  21. [22]

    2024 , eprint=

    Cambrian-1: A Fully Open, Vision-Centric Exploration of Multimodal LLMs , author=. 2024 , eprint=

    2024

  22. [23]

    2024 , eprint=

    Are We on the Right Way for Evaluating Large Vision-Language Models? , author=. 2024 , eprint=

    2024

  23. [25]

    MiniLLM: Knowledge Distillation of Large Language Models , url =

    Gu, Yuxian and Dong, Li and Wei, Furu and Huang, Minlie , booktitle =. MiniLLM: Knowledge Distillation of Large Language Models , url =

  24. [26]

    2024 , editor =

    Ko, Jongwoo and Kim, Sungnyun and Chen, Tianyi and Yun, Se-Young , booktitle =. 2024 , editor =

    2024

  25. [27]

    2025 , editor =

    Ko, Jongwoo and Chen, Tianyi and Kim, Sungnyun and Ding, Tianyu and Liang, Luming and Zharkov, Ilya and Yun, Se-Young , booktitle =. 2025 , editor =

    2025

  26. [28]

    2026 , eprint =

    On-Policy Context Distillation for Language Models , author =. 2026 , eprint =

    2026

  27. [29]

    2024 , eprint =

    LLaVA-UHD: an LMM Perceiving Any Aspect Ratio and High-Resolution Images , author =. 2024 , eprint =

    2024

  28. [30]

    2024 , eprint =

    Mini-Gemini: Mining the Potential of Multi-modality Vision Language Models , author =. 2024 , eprint =

    2024

  29. [31]

    Z oom E ye: Enhancing Multimodal LLM s with Human-Like Zooming Capabilities through Tree-Based Image Exploration

    Shen, Haozhan and Zhao, Kangjia and Zhao, Tiancheng and Xu, Ruochen and Zhang, Zilun and Zhu, Mingwei and Yin, Jianwei. Z oom E ye: Enhancing Multimodal LLM s with Human-Like Zooming Capabilities through Tree-Based Image Exploration. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.335

    doi:10.18653/v1/2025.emnlp-main.335 2025

  30. [32]

    2025 , eprint =

    ReFocus: Visual Editing as a Chain of Thought for Structured Image Understanding , author =. 2025 , eprint =

    2025

  31. [33]

    Thinking with Images

    DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning , author =. 2026 , eprint =

    2026

  32. [34]

    2025 , eprint =

    Thyme: Think Beyond Images , author =. 2025 , eprint =

    2025

  33. [35]

    2026 , eprint =

    DeepEyesV2: Toward Agentic Multimodal Model , author =. 2026 , eprint =

    2026

  34. [36]

    2026 , eprint =

    SenseNova-MARS: Empowering Multimodal Agentic Reasoning and Search via Reinforcement Learning , author =. 2026 , eprint =

    2026

  35. [37]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =

    Leng, Sicong and Zhang, Hang and Chen, Guanzheng and Li, Xin and Lu, Shijian and Miao, Chunyan and Bing, Lidong , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2024 , pages =

    2024

  36. [38]

    2024 , editor =

    Chen, Zhaorun and Zhao, Zhuokai and Luo, Hongyin and Yao, Huaxiu and Li, Bo and Zhou, Jiawei , booktitle =. 2024 , editor =

    2024

  37. [39]

    2024 , eprint =

    Delve into Visual Contrastive Decoding for Hallucination Mitigation of Large Vision-Language Models , author =. 2024 , eprint =

    2024

  38. [40]

    2026 , eprint =

    Thinking with Deltas: Incentivizing Reinforcement Learning via Differential Visual Reasoning Policy , author =. 2026 , eprint =

    2026

  39. [41]

    2026 , eprint =

    Visual-OPSD: Cross-Modal On-Policy Self-Distillation for Efficient Unified Multimodal Reasoning , author =. 2026 , eprint =

    2026

  40. [42]

    2026 , eprint =

    Med-OPD: Improving Medical Vision-Language Models via Evidence-Aware On-Policy Distillation , author =. 2026 , eprint =

    2026

  41. [43]

    2026 , eprint =

    Contrastive On-Policy Distillation , author =. 2026 , eprint =

    2026

  42. [46]

    2025 , month = dec, url =

    2025

  43. [47]

    2026 , month = feb, url =

    2026

  44. [48]

    2025 , month = dec, url =

    Introducing. 2025 , month = dec, url =

    2025

  45. [49]

    2026 , month = mar, url =

    Introducing. 2026 , month = mar, url =

    2026

  46. [50]

    doi:10.48550/arXiv.2508.10925 , url =

    2025 , eprint =. doi:10.48550/arXiv.2508.10925 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2508.10925 2025

  47. [51]

    On-policy distillation of language models: Learning from self-generated mistakes, 2024

    Rishabh Agarwal, Nino Vieillard, Yongchao Zhou, Piotr Stanczyk, Sabela Ramos, Matthieu Geist, and Olivier Bachem. On-policy distillation of language models: Learning from self-generated mistakes, 2024. URL https://arxiv.org/abs/2306.13649

    Pith/arXiv arXiv 2024

  48. [52]

    Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. URL https://arxiv.org/abs/2308.12966

    Pith/arXiv arXiv 2023

  49. [53]

    Qwen3-VL technical report, 2025 a

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, Wenbin Ge, Zhifang Guo, Qidong Huang, Jie Huang, Fei Huang, Binyuan Hui, Shutong Jiang, Zhaohai Li, Mingsheng Li, Mei Li, Kaixin Li, Zicheng Lin, Junyang Lin, Xuejing Liu, Jiawei Liu, Chenglong Liu, Yang Liu, Dayiheng Liu, Shixuan ...

    Pith/arXiv arXiv 2025

  50. [54]

    Qwen2.5-vl technical report, 2025 b

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report, 2...

    Pith/arXiv arXiv 2025

  51. [55]

    Are we on the right way for evaluating large vision-language models?, 2024 a

    Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Zehui Chen, Haodong Duan, Jiaqi Wang, Yu Qiao, Dahua Lin, and Feng Zhao. Are we on the right way for evaluating large vision-language models?, 2024 a . URL https://arxiv.org/abs/2403.20330

    Pith/arXiv arXiv 2024

  52. [56]

    HALC : Object hallucination reduction via adaptive focal-contrast decoding

    Zhaorun Chen, Zhuokai Zhao, Hongyin Luo, Huaxiu Yao, Bo Li, and Jiawei Zhou. HALC : Object hallucination reduction via adaptive focal-contrast decoding. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning,...

    2024

  53. [57]

    Sensenova-mars: Empowering multimodal agentic reasoning and search via reinforcement learning, 2026

    Yong Xien Chng, Tao Hu, Wenwen Tong, Xueheng Li, Jiandong Chen, Haojia Yu, Jiefan Lu, Hewei Guo, Hanming Deng, Chengjun Xie, Gao Huang, Dahua Lin, and Lewei Lu. Sensenova-mars: Empowering multimodal agentic reasoning and search via reinforcement learning, 2026. URL https://arxiv.org/abs/2512.24330

    arXiv 2026

  54. [58]

    Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023

    Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, and Steven Hoi. Instructblip: Towards general-purpose vision-language models with instruction tuning, 2023. URL https://arxiv.org/abs/2305.06500

    Pith/arXiv arXiv 2023

  55. [59]

    Refocus: Visual editing as a chain of thought for structured image understanding, 2025

    Xingyu Fu, Minqian Liu, Zhengyuan Yang, John Corring, Yijuan Lu, Jianwei Yang, Dan Roth, Dinei Florencio, and Cha Zhang. Refocus: Visual editing as a chain of thought for structured image understanding, 2025. URL https://arxiv.org/abs/2501.05452

    Pith/arXiv arXiv 2025

  56. [60]

    Thinking with deltas: Incentivizing reinforcement learning via differential visual reasoning policy, 2026

    Shujian Gao, Yuan Wang, Jiangtao Yan, Zuxuan Wu, and Yu-Gang Jiang. Thinking with deltas: Incentivizing reinforcement learning via differential visual reasoning policy, 2026. URL https://arxiv.org/abs/2601.06801

    arXiv 2026

  57. [61]

    Gemini 3 Flash : Model card, December 2025

    Google DeepMind . Gemini 3 Flash : Model card, December 2025. URL https://deepmind.google/models/model-cards/gemini-3-flash/. Published December 17, 2025

    2025

  58. [62]

    Gemini 3.1 Pro : Model card, February 2026

    Google DeepMind . Gemini 3.1 Pro : Model card, February 2026. URL https://deepmind.google/models/model-cards/gemini-3-1-pro/. Published February 19, 2026

    2026

  59. [63]

    Minillm: Knowledge distillation of large language models

    Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In B. Kim, Y. Yue, S. Chaudhuri, K. Fragkiadaki, M. Khan, and Y. Sun, editors, International Conference on Learning Representations, volume 2024, pages 32694--32717, 2024. URL https://proceedings.iclr.cc/paper_files/paper/2024/file/8ac015d409635f196f9...

    2024

  60. [64]

    Distilling the knowledge in a neural network, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. URL https://arxiv.org/abs/1503.02531

    Pith/arXiv arXiv 2015

  61. [65]

    Deepeyesv2: Toward agentic multimodal model, 2026

    Jack Hong, Chenxiao Zhao, ChengLin Zhu, Weiheng Lu, Guohai Xu, and Xing Yu. Deepeyesv2: Toward agentic multimodal model, 2026. URL https://arxiv.org/abs/2511.05271

    Pith/arXiv arXiv 2026

  62. [66]

    Yoon Kim and Alexander M. Rush. Sequence-level knowledge distillation. In Jian Su, Kevin Duh, and Xavier Carreras, editors, Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1317--1327, Austin, Texas, November 2016. Association for Computational Linguistics. doi:10.18653/v1/D16-1139. URL https://aclanthology.org...

    doi:10.18653/v1/d16-1139 2016

  63. [67]

    D isti LLM : Towards streamlined distillation for large language models

    Jongwoo Ko, Sungnyun Kim, Tianyi Chen, and Se-Young Yun. D isti LLM : Towards streamlined distillation for large language models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors, Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedi...

    2024

  64. [68]

    D isti LLM -2: A contrastive approach boosts the distillation of LLM s

    Jongwoo Ko, Tianyi Chen, Sungnyun Kim, Tianyu Ding, Luming Liang, Ilya Zharkov, and Se-Young Yun. D isti LLM -2: A contrastive approach boosts the distillation of LLM s. In Aarti Singh, Maryam Fazel, Daniel Hsu, Simon Lacoste-Julien, Felix Berkenkamp, Tegan Maharaj, Kiri Wagstaff, and Jerry Zhu, editors, Proceedings of the 42nd International Conference on...

    2025

  65. [69]

    Delve into visual contrastive decoding for hallucination mitigation of large vision-language models, 2024

    Yi-Lun Lee, Yi-Hsuan Tsai, and Wei-Chen Chiu. Delve into visual contrastive decoding for hallucination mitigation of large vision-language models, 2024. URL https://arxiv.org/abs/2412.06775

    Pith/arXiv arXiv 2024

  66. [70]

    Mitigating object hallucinations in large vision-language models through visual contrastive decoding

    Sicong Leng, Hang Zhang, Guanzheng Chen, Xin Li, Shijian Lu, Chunyan Miao, and Lidong Bing. Mitigating object hallucinations in large vision-language models through visual contrastive decoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13872--13882, June 2024

    2024

  67. [71]

    Visual-opsd: Cross-modal on-policy self-distillation for efficient unified multimodal reasoning, 2026

    Pengyu Li, Zhitao Gao, Lingling Zhang, Muye Huang, Yuanming Li, Fangzhi Xu, and Jun Liu. Visual-opsd: Cross-modal on-policy self-distillation for efficient unified multimodal reasoning, 2026. URL https://arxiv.org/abs/2606.18974

    Pith/arXiv arXiv 2026

  68. [72]

    Mini-gemini: Mining the potential of multi-modality vision language models, 2024

    Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models, 2024. URL https://arxiv.org/abs/2403.18814

    Pith/arXiv arXiv 2024

  69. [73]

    Evaluating object hallucination in large vision-language models

    Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. In Houda Bouamor, Juan Pino, and Kalika Bali, editors, Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 292--305, Singapore, December 2023. Association for Computational Lingui...

    doi:10.18653/v1/2023.emnlp-main.20 2023

  70. [74]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. URL https://arxiv.org/abs/2304.08485

    Pith/arXiv arXiv 2023

  71. [75]

    Visual-advantage on-policy distillation for vision-language models, 2026

    Ruiqi Liu, Xiaolei Lv, Gengsheng Li, Ximo Zhu, Zhiheng Wang, Zhengbo Zhang, Junkai Chen, Zhiheng Li, Bo Li, Jun Gao, and Shu Wu. Visual-advantage on-policy distillation for vision-language models, 2026. URL https://arxiv.org/abs/2605.21924

    Pith/arXiv arXiv 2026

  72. [76]

    Kimi K2.6 : From code to creation, from one to many, 2026

    Moonshot AI . Kimi K2.6 : From code to creation, from one to many, 2026. URL https://www.kimi.com/ai-models/kimi-k2-6. Official model page

    2026

  73. [77]

    Introducing GPT-5.2 , December 2025 a

    OpenAI . Introducing GPT-5.2 , December 2025 a . URL https://openai.com/index/introducing-gpt-5-2/. Published December 11, 2025

    2025

  74. [78]

    gpt-oss-120b & gpt-oss-20b model card, 2025 b

    OpenAI . gpt-oss-120b & gpt-oss-20b model card, 2025 b . URL https://arxiv.org/abs/2508.10925

    Pith/arXiv arXiv 2025

  75. [79]

    Introducing GPT-5.4 , March 2026

    OpenAI . Introducing GPT-5.4 , March 2026. URL https://openai.com/index/introducing-gpt-5-4/. Published March 5, 2026

    2026

  76. [80]

    Med-opd: Improving medical vision-language models via evidence-aware on-policy distillation, 2026

    Yunhang Qian, Jiaquan Yu, Jiawei Liu, Meng Wang, Hongwei Bran Li, and Xiaobin Hu. Med-opd: Improving medical vision-language models via evidence-aware on-policy distillation, 2026. URL https://arxiv.org/abs/2607.16303

    Pith/arXiv arXiv 2026

  77. [81]

    Qwen3.5 : Towards native multimodal agents, 2026

    Qwen Team . Qwen3.5 : Towards native multimodal agents, 2026. URL https://qwen.ai/blog?id=qwen3.5. Official model release and technical blog

    2026

  78. [82]

    Contrastive on-policy distillation, 2026

    Jiacheng Ruan, Jun Tang, Wenzhen Yuan, Ting Liu, Shuai Bai, Dayiheng Liu, Zhibo Yang, and Yuzhuo Fu. Contrastive on-policy distillation, 2026. URL https://arxiv.org/abs/2607.19046

    Pith/arXiv arXiv 2026

  79. [83]

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

    Pith/arXiv arXiv 2024

  80. [84]

    Z oom E ye: Enhancing multimodal LLM s with human-like zooming capabilities through tree-based image exploration

    Haozhan Shen, Kangjia Zhao, Tiancheng Zhao, Ruochen Xu, Zilun Zhang, Mingwei Zhu, and Jianwei Yin. Z oom E ye: Enhancing multimodal LLM s with human-like zooming capabilities through tree-based image exploration. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors, Proceedings of the 2025 Conference on Empirical Meth...

    2025

Showing first 80 references.