Pith. sign in

REVIEW 2 major objections 5 minor 45 references

Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots

T0 review · 2 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A model learns to see better from its own counterfactual blind spots.

desk verdict CVPD's counterfactual blind-spot criterion is a real idea, but the undisclosed 15k-image training pool makes the headline benchmark gains uncheckable until contamination is ruled out. read the letter →

arxiv 2608.09931 v1 pith:4JYA3CNB submitted 2026-08-10 cs.CV

Shravan Venkatraman , Omkar Thawakar , Ritesh Thawkar , Abdelrahman Shaker , Rao Muhammad Anwer This is my paper · ORCID
classification cs.CV
keywords visualself-distillationmultimodallargelanguagemodelscounterfactualblindspotson-policydistillationself-improvementcontrastivelearningfine-grainedperceptionOCRBench
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 claims that a multimodal large language model can improve its own fine-grained visual perception using no external labels, tools, rewards, or stronger teachers. It locates regions where cropping the image changes and sharpens the model's answer distribution, while blurring that same region leaves the full-image behavior almost unchanged; such regions are diagnosed as 'blind spots' where the model has perceptual ability it fails to use under full-image conditioning. CVPD turns each discovered region into dense token-level supervision, training the full-image student toward its own crop-conditioned answers and away from its ghost-conditioned default, with a KL anchor preserving general capabilities. On the 8-billion-parameter instruct model this yields gains of +3.60 on OCRBench, +3.38 on MMStar Fine-Grained Perception, and +3.08 on MMStar Logical Reasoning, with no regressions across twelve benchmarks.

What carries the argument

The central machinery is the three-gate Counterfactual Criterion (Equations 2–4) together with the contrastive self-distillation objective. A region $R$ is a blind spot when the Jensen–Shannon divergence between crop-conditioned and full-image distributions exceeds a threshold, the divergence between ghost-conditioned and full-image distributions stays below a threshold, and the crop-conditioned prediction has lower entropy than the full-image prediction. That selection step converts raw unlabeled images into a curated pool, and the training loss then makes the crop-conditioned distribution the positive teacher and the ghost-conditioned distribution the negative teacher at every token position, while a KL anchor to the frozen reference policy limits drift in general capabilities.

What would settle it

On a labeled subset of images, compare answer accuracy under the full image, the crop, and the ghost for regions that pass all three gates. If crop-conditioned answers are not systematically more accurate than full-image answers on these regions, Gate 3 is selecting confidence rather than capability and the central mechanism fails; if they are more accurate, the paper's premise holds.

Watch

Extended reading notes

Core claim

CVPD's central assertion is that the three-gate Counterfactual Criterion identifies regions where the crop view and the ghost view are maximally opposed: the crop must diverge from the full image, the ghost must stay close to it, and the crop must lower predictive entropy. Regions passing all three gates are interpreted as visual blind spots where the model encodes information it does not consistently exploit under full-image conditioning, and the ghost view is a faithful proxy for its inattentive default. These regions form a curated pool of (image, question, region, answer) tuples, and training instantiates four policies from the same backbone: an online full-image student, a momentum crop-conditioned teacher, a momentum ghost-conditioned teacher, and a frozen reference policy. The combined loss transfers the crop teacher into the student, ranks the crop teacher above the ghost teacher with a margin, and anchors the student to the reference policy. On Qwen3-VL-8B-Instruct the method outperforms six self-evolving baselines on all twelve benchmarks without a single regression, including baselines that rely on an external annotator.

Load-bearing premise

The load-bearing premise is that a crop that makes the model more confident is revealing latent correct perception, since nothing in the pipeline checks that the sharper answer is actually the right one.

Editorial extensions

If this is right

  • Dense token-level visual supervision can be recovered from a model's own counterfactual responses, eliminating the need for external region proposals, segmentation systems, rewards, or stronger models.
  • The largest gains appear exactly on tasks requiring localized attention: OCRBench (+3.60), MMStar Fine-Grained Perception (+3.38), and MMStar Logical Reasoning (+3.08) at the 8B scale.
  • The contrastive ghost teacher is an active training signal, not just a filter: removing the ranking objective drops OCRBench by 2.30 points and MMStar Fine-Grained Perception by 2.16 points.
  • The Counterfactual Criterion is the main driver of improvement: replacing it with random regions causes the largest ablation drop, 2.60 points on OCRBench and 2.53 on MMStar Fine-Grained Perception.
  • The same pipeline transfers to the 4-billion-parameter model, improving every benchmark at both scales with no regression.

Reading between the lines

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

  • A natural extension the paper leaves implicit is an iterative loop: re-running blind-spot discovery on the trained student should surface harder regions, and one could test whether gains compound or plateau across multiple rounds.
  • Because the criterion uses only distributional divergence and entropy, it should transfer to other token-modalities: for example, treating a temporal window in audio or a patch in video as the 'crop' and time-masking or patch-blurring as the 'ghost.'
  • The method's premise that entropy sharpening equals perceptual correctness can be tested directly: label a sample of curated tuples and compare crop-conditioned accuracy against full-image accuracy; the framework could then be strengthened by a confidence-weighting scheme that down-weights confidently wrong crops.
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

2 major / 5 minor

Summary. The paper introduces CVPD, a self-distillation method for multimodal large language models that derives dense token-level supervision from the model's own counterfactual responses on unlabeled images. A three-gate Counterfactual Criterion identifies regions where a cropped view sharpens the model's answer distribution while a ghosted (blurred) view leaves the full-image distribution largely unchanged. Phase 2 then trains a LoRA student using the crop-conditioned EMA teacher as a positive target and the ghost-conditioned EMA teacher as a negative target, together with a KL anchor to the frozen reference policy. Experiments on Qwen3-VL-4B and 8B claim consistent gains over six self-evolving baselines across twelve benchmarks, with the largest improvements on OCRBench and MMStar fine-grained and logical reasoning subscales. Ablations show that replacing the counterfactual criterion with random regions and removing the contrastive ranking objective both cause substantial drops, and hyperparameter sensitivity for the ranking weight, margin, and KL target is reported as flat.

Significance. If the result holds, CVPD is a meaningful step toward fully self-contained visual self-distillation: it removes the need for external region annotations, segmentation tools, rewards, or stronger teachers, and it provides per-token rather than scalar supervision. The counterfactual blind-spot criterion is a novel operationalization of 'latent perceptual capacity,' and the paper supports it with a focused ablation study, a blind-spot source analysis, and a sensitivity analysis. The paper also releases code, a model, and a project page, which strengthens reproducibility once the data provenance is clarified. The main open question is whether the benchmark gains reflect genuine distillation or overlap between the undisclosed unlabeled pool and the evaluation sets, and whether the crop-conditioned teacher is trustworthy in the absence of any correctness check.

major comments (2)
  1. [5 Training Data / Table 1] The central empirical claim, including the headline +3.60 on OCRBench, +3.38 on MMStar Fine-Grained Perception, and the 'without a single regression' statement, presupposes that the 15,000-image unlabeled pool is disjoint from the evaluation benchmarks. The Training Data section does not name the image sources, provide hashes, or describe any deduplication; it only says that Stage 2 consists of charts, scientific diagrams, and structured document figures plus natural scenes. The evaluation suite includes AI2D, ScienceQA, InfoVQA, OCRBench, CV-Bench, and RealWorldQA, which contain exactly these modalities. Because the model is trained with LoRA on these images, overlapping images could inflate per-image recall, especially for OCR strings and layout patterns, without improving perception. This is load-bearing, not a reproducibility nicety, and it is fixable: disclose sources and hashes, run exact and near-duplicate deduplication against all evaluation sets, and report results on a provably disjoint subset.
  2. [4.1 Eq. (4), 4.2 Eq. (6)] Gate G3 equates a drop in predictive entropy under the crop with epistemic sharpening, but without labels or an external verifier there is no check that the sharpened distribution is more correct rather than merely more confident. A crop that confidently produces a wrong answer passes all three gates and is then reinforced by the latent-transfer and contrastive-ranking objectives in Eq. (6). Figure 3 reports distributional statistics only (divergences and entropy deltas), not crop-conditioned accuracy, so the claim that these regions 'reveal perceptual information the model can encode' is not directly evidenced. The manuscript should provide a labeled audit subset, a human evaluation, or a cross-view consistency check that verifies crop-conditioned answers are at least as accurate as full-image answers before treating the crop view as a trustworthy teacher.
minor comments (5)
  1. [4.1 Eq. (4)] Gate G3 uses only the first token y0 for the entropy comparison, while the distillation objective in Eq. (6) averages over all rollout tokens; please justify this choice or extend the criterion to the full probe answer.
  2. [Algorithm 1 / Table 4] The track membership percentages sum to 110% (27% + 53% + 30%); the text says this is due to multi-region retention, but reporting per-track counts or a Venn-style breakdown would make the numbers easier to interpret.
  3. [Table 2 row (d)] The 'no curation (random regions)' ablation should specify how the random regions are generated; if they are uniformly random boxes within the same 1%-50% area range, the comparison to the gated criterion is cleaner than if they are sampled from a different distribution.
  4. [5.3 Sensitivity Analysis] The sensitivity analysis covers lambda_rank, margin m, and KL target kappa, but not the discovery thresholds tau_crop and tau_ghost; since these thresholds control the 17.2% yield, a brief sweep would strengthen the claim that the criterion is not finely tuned.
  5. [5 Training Data] The sentence 'with no captions, bounding boxes, semantic labels, ground-truth answers, reward signals, or external models are used at any stage' contains a grammar error and should be reworded, for example as 'no captions, bounding boxes, semantic labels, ground-truth answers, reward signals, or external models are used at any stage.'

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity in the auxiliary validation of the Counterfactual Criterion; the central benchmark claims remain externally grounded.

  1. self definitional [Section 4.1, Figure 3 analysis; echoed in Section 5.2]
    "The crop-versus-ghost gap rises from −0.028 to +0.249 after curation, supporting the view that the criterion isolates regions where the two teachers provide genuine and opposing supervision."

    Equations (2)–(4) define a blind spot directly in terms of the Figure 3 statistics: D_JS(crop||full) ≥ τ_crop, D_JS(ghost||full) ≤ τ_ghost, and H[p_crop] < H[p_full]. Any region passing the gates must therefore, by construction, have a large crop-vs-ghost divergence gap and a positive entropy delta; the reported 'after curation' numbers are arithmetic consequences of the selection rule, not independent evidence that the selected regions provide 'genuine' supervision. The criterion's value is instead established by the external benchmark ablation (row d of Table 2), which compares curated versus random regions on held-out tasks. That external comparison is not circular, so this self-referential evidence step does not undermine the paper's central empirical claim.

full rationale

The main derivation chain is not circular: CVPD discovers regions from the model's own crop and ghost responses, constructs paired teacher distributions from the same backbone, distills them into the full-image student, and evaluates on twelve external benchmarks. Those benchmarks are independent of the training responses, so the reported gains are not identical by construction to the distillation signal. The undisclosed 15,000-image unlabeled pool and the absence of deduplication against evaluation benchmarks is a contamination and external-validity concern, not a circularity of the derivation. Self-citations to EvoLMM and related work are used as baselines or prior art rather than as load-bearing justification for the method's core premise. The only circular element found is the Figure 3 validation, which reports exactly the quantities the three gates are defined to maximize or minimize; that step is auxiliary, and the external ablation study provides independent support for the criterion. Overall, the paper's predictions are externally measured rather than fitted inputs, so the circularity score is low.

Assumptions & free parameters 9 free parameters · 5 assumptions · 1 invented entities

The method is a self-distillation recipe whose supervision is generated from the model's own distributions, so the conceptual labels are internally defined. The central performance claim is externally anchored by twelve benchmarks, keeping the circularity burden low, but the free parameters are hand-set or tuned on the evaluation suite, and the assumption that sharper crops are more correct is unverified.

free parameters (9)
  • tau_crop (G1 divergence threshold) = 0.05
    Hand-set threshold for crop/full JS divergence in Eq. (2); determines which regions count as blind spots. No sensitivity study is reported for this threshold.
  • tau_ghost (G2 divergence threshold) = 0.05
    Hand-set threshold for ghost/full invariance in Eq. (3); same value as tau_crop. Part of the three-gate selection criterion.
  • lambda_rank (contrastive ranking weight) = 0.5
    Weight on the contrastive ranking objective in Eq. (6); selected via sensitivity analysis on evaluation benchmarks in Section 5.3.
  • margin m = 0.1
    Margin in the contrastive ranking loss in Eq. (6); tuned through the sensitivity table.
  • KL target kappa = 0.03
    Target KL divergence for the anchor regularization; adapted online and selected via sensitivity analysis.
  • EMA coefficient alpha = 0.05
    Momentum teacher update coefficient in Algorithm 1 and Section 4.2; ablated in Table 2.
  • Crop margin and upscaling = 20% margin, upscaled to original resolution
    Defines the crop view in Section 4.1; not swept, but affects which regions pass the criterion.
  • Candidate area filter = 1% to 50% of image area
    Candidates outside this range are discarded as degenerate in Section 4.2.
  • Top-K for JS approximation = 100
    Per-token D_JS is approximated over the union of top-K logit indices in Section 3.
assumptions (5)
  • domain assumption A reduction in per-token entropy under the crop view (G3) indicates a more accurate answer distribution, not merely a more confident one.
    Eq. (4) uses H[p_crop(y0)] < H[p_full(y0)] as 'Epistemic Sharpening' with no correctness labels. If the sharpened answer is wrong, distillation would reinforce errors.
  • domain assumption Gaussian blur of a region leaves the full-image distribution unchanged (G2) iff the model is not using that region, and the ghost view is a faithful proxy for the model's inattentive default.
    Eq. (3) and Section 4.2 treat p_ghost as a counterfactual negative teacher. Blurring can also produce unrecognized artifacts rather than pure removal of information.
  • domain assumption The top-K union JS approximation over one short probe answer is a reliable estimate of the true per-token distributional gap between views.
    Section 3 defines D = D_JS with K=100 over a fixed teacher-forced probe answer. This is an approximation of full next-token distributions.
  • domain assumption Self-generated questions and probe answers expose the fine-grained perception behaviors that transfer to external benchmarks.
    Phase 1 generates q and a with the model itself; the method assumes these self-generated questions cover the localized perception skills evaluated by OCRBench, MMStar, InfoVQA, and similar benchmarks.
  • domain assumption Training on unlabeled images from the same broad domains as evaluation benchmarks does not create misleading distribution overlap.
    Section 5 uses 10,000 natural-scene images plus 5,000 reasoning-domain images. Sources are not disclosed, so overlap with AI2D, ScienceQA, InfoVQA, and OCRBench cannot be ruled out.
invented entities (1)
  • Counterfactual blind spot
    purpose: Operational construct for selecting regions that provide opposing crop and ghost supervision in Phase 1.
    A 'visual blind spot' is defined entirely through the model's own response statistics in Eqs. (2)-(4) and is validated only through downstream benchmark gains. There is no independent, externally labeled confirmation that selected regions are actual perception errors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots." pith.science (2026). https://pith.science/paper/4JYA3CNB

@misc{pith2026260809931,
  author       = {Pith},
  title        = {Pith review of: Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4JYA3CNB}},
  note         = {Machine review of arXiv:2608.09931}
}
abstract

Self-improvement for multimodal large language models (MLLMs) is typically driven by reward-based methods that provide only coarse scalar feedback. Distillation offers a richer alternative through dense token-level supervision, but in the visual domain it usually depends on privileged context constructed using external annotations and tools, or stronger models. We introduce \textbf{CVPD} (Contrastive Counterfactual Visual Process Distillation), which, to the best of our knowledge, is the first fully self-contained framework for dense, on-policy, token-level visual self-distillation for MLLMs. CVPD identifies visual blind spots where zooming into a region changes and sharpens the model's answer distribution, while removing the same region leaves the full-image behavior largely unchanged. Such regions reveal perceptual information that the model can encode but fails to consistently utilize under full-image conditioning. We propose a three-gate Counterfactual Criterion that identifies these regions directly from the model's own responses and converts them into dense contrastive supervision for self-distillation. On Qwen3-VL-8B-Instruct, CVPD outperforms six self-evolving baselines across twelve benchmarks, including methods that rely on external GPT-4o supervision, without a single regression. It achieves gains of $+3.60$ on OCRBench, $+3.38$ on MMStar Fine-Grained Perception, and $+3.08$ on MMStar Logical Reasoning, while maintaining or improving performance on broader multimodal benchmarks.

Figures

Figures reproduced from arXiv: 2608.09931 by the authors.

Figure 1
Figure 1. CVPD turns visual blind spots into dense self-supervision. Reward-based self-evolution can identify better or worse responses, but it does not provide token-level corrective distributions. Standard visual distillation can provide dense supervision, but typically relies on externally constructed privileged views. CVPD instead uses the model’s own responses to identify regions where a crop sharpens the answer distribu… view at source ↗
Figure 2
Figure 2. Overview of CVPD. CVPD derives all supervision from the model’s own responses over unlabeled images, without external labels, annotation tools, segmentation models, or stronger annotators. Phase 1 (Counterfactual Blind-Spot Discovery): The model generates a fine-grained question–answer pair (q, a) for each image and proposes candidate regions through self-grounding (Track A), a 3×3 grid (Track B), and a 2×2 grid (Tr… view at source ↗
Figure 3
Figure 3. Distributional signature of counterfactual blind-spot curation. We sample 6,791 candidate regions from 500 raw images. (a) Uncurated regions produce similar crop/full and ghost/full divergences (0.359 vs. 0.387), indicating that both interventions perturb the model in comparable ways. Regions passing the Counterfactual Crite￾rion retain strong crop-side divergence (0.265) while reducing ghost-side divergence to 0.01… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 13 canonical work pages

  1. [1]

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

    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. InInternational Conference on Learning Representations, 2024

    work page 2024

  2. [2]

    Qwen3-VL technical report, 2025

    Shuai Bai et al. Qwen3-VL technical report, 2025. arXiv preprint arXiv:2511.21631

    arXiv 2025

  3. [3]

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

    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?,

    work page

  4. [4]

    Mme: A comprehensive evaluation benchmark for multimodal large language models

    Chaoyou Fu et al. Mme: A comprehensive evaluation benchmark for multimodal large language models. Advances in Neural Information Processing Systems, 38, 2026

    work page 2026

  5. [5]

    Reinforced self-training (ReST) for language modeling, 2023

    Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, et al. Reinforced self-training (ReST) for language modeling, 2023. arXiv preprint arXiv:2308.08998. 11

    arXiv 2023

  6. [6]

    Active-Zero: Self-evolving vision-language models through active environment exploration, 2026

    Jinghan He, Junfeng Fang, Feng Xiong, Zijun Yao, Fei Shen, Haiyun Guo, Jinqiao Wang, and Tat-Seng Chua. Active-Zero: Self-evolving vision-language models through active environment exploration, 2026. arXiv preprint arXiv:2602.11241

    arXiv 2026

  7. [7]

    VisPlay: Self-evolving vision-language models from images, 2025

    Yicheng He, Chengsong Huang, Zongxia Li, Jiaxin Huang, and Yonghui Yang. VisPlay: Self-evolving vision-language models from images, 2025. arXiv preprint arXiv:2511.15661

    arXiv 2025

  8. [8]

    CEPO: RLVR Self-Distillation using Contrastive Evidence Policy Optimization

    Ahmed Heakl, Abdelrahman M. Shaker, Youssef Mohamed, Rania Elbadry, Omar Fetouh, Fahad Shahbaz Khan, and Salman Khan. Cepo: Rlvr self-distillation using contrastive evidence policy optimization.arXiv preprint arXiv:2605.19436, 2026

    work page Pith review arXiv 2026

Show all 45 references
  1. [9]

    Distilling the knowledge in a neural network, 2015

    Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network, 2015. arXiv preprint arXiv:1503.02531

    2015 arXiv

  2. [10]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InInternational Conference on Learning Representations, 2022

    2022

  3. [11]

    Reinforcement learning via self-distillation, 2026

    Jonas Hübotter, Frederike Lübeck, Lejs Behric, et al. Reinforcement learning via self-distillation, 2026. arXiv preprint arXiv:2601.20802

    2026 arXiv

  4. [12]

    A diagram is worth a dozen images

    Aniruddha Kembhavi, Mike Salvato, Eric Kolve, Minjoon Seo, Hannaneh Hajishirzi, and Ali Farhadi. A diagram is worth a dozen images. InEuropean conference on computer vision, pages 235–251. Springer, 2016

    2016

  5. [13]

    Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125, 2023

    Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension.arXiv preprint arXiv:2307.16125, 2023

    2023 arXiv

  6. [14]

    MM-Zero: Self-evolving multi-model vision language models from zero data, 2026

    Zongxia Li, Hongyang Du, Chengsong Huang, Xiyang Wu, Lantao Yu, Yicheng He, Jing Xie, Xiaomin Wu, Zhichao Liu, Jiarui Zhang, and Fuxiao Liu. MM-Zero: Self-evolving multi-model vision language models from zero data, 2026. arXiv preprint arXiv:2603.09206

    2026

  7. [15]

    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. arXiv preprint arXiv:2605.21924

    2026 arXiv

  8. [16]

    Mmbench: Is your multi-modal model an all-around player? InEuropean conference on computer vision

    Yuan Liu et al. Mmbench: Is your multi-modal model an all-around player? InEuropean conference on computer vision. Springer, 2024

    2024

  9. [17]

    Ocrbench: on the hidden mystery of ocr in large multimodal models.Science China Information Sciences, 67(12), December 2024

    Yuliang Liu, Zhang Li, Mingxin Huang, Biao Yang, Wenwen Yu, Chunyuan Li, Xu-Cheng Yin, Cheng-Lin Liu, Lianwen Jin, and Xiang Bai. Ocrbench: on the hidden mystery of ocr in large multimodal models.Science China Information Sciences, 67(12), December 2024

    2024

  10. [18]

    Decoupled weight decay regularization, 2019

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019

    2019

  11. [19]

    Learn to explain: Multimodal reasoning via thought chains for science question answering, 2022

    Pan Lu, Swaroop Mishra, Tony Xia, Liang Qiu, Kai-Wei Chang, Song-Chun Zhu, Oyvind Tafjord, Peter Clark, and Ashwin Kalyan. Learn to explain: Multimodal reasoning via thought chains for science question answering, 2022

    2022

  12. [20]

    Minesh Mathew, Viraj Bagal, Rubèn Tito, Dimosthenis Karatzas, Ernest Valveny, and C.V . Jawahar. Info- graphicvqa. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1697–1706, 2022

    2022

  13. [21]

    Gpt-4 technical report, 2024

    OpenAI et al. Gpt-4 technical report, 2024

    2024

  14. [22]

    A reduction of imitation learning and structured prediction to no-regret online learning

    Stephane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. InProceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, volume 15 ofProceedings of Machine Le...

    2011

  15. [23]

    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. arXiv preprint arXiv:2402.03300

    2024 arXiv

  16. [24]

    VLM-R1: A stable and generalizable R1-style large vision-language model, 2025

    Haozhan Shen, Peng Liu, Jingcheng Li, Chunxin Fang, Yibo Ma, Jiajia Liao, Qiaoli Shen, Zilun Zhang, Kangjia Zhao, Qianqian Zhang, Ruochen Xu, and Tiancheng Zhao. VLM-R1: A stable and generalizable R1-style large vision-language model, 2025. arXiv preprint arXiv:2504.07615. 12

    2025 arXiv

  17. [25]

    iReasoner: Trajectory- aware intrinsic reasoning supervision for self-evolving large multimodal models, 2026

    Meghana Sunil, Manikandarajan Venmathimaran, and Muthu Subash Kavitha. iReasoner: Trajectory- aware intrinsic reasoning supervision for self-evolving large multimodal models, 2026. arXiv preprint arXiv:2601.05877

    2026 arXiv

  18. [26]

    EvoLMM: Self-evolving large multimodal models with continuous rewards, 2026

    Omkar Thawakar, Shravan Venkatraman, Ritesh Thawkar, Abdelrahman Shaker, Hisham Cholakkal, Rao Muhammad Anwer, Salman Khan, and Fahad Khan. EvoLMM: Self-evolving large multimodal models with continuous rewards, 2026. arXiv preprint arXiv:2511.16672

    2026 arXiv

  19. [27]

    Ask, solve, generate: Self-evolving unified multimodal understanding and generation via self-consistency rewards.arXiv preprint arXiv:2606.27376, 2026

    Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar, Abdelrahman Shaker, Fahad Khan, Hisham Cholakkal, Salman Khan, and Rao Muhammad Anwer. Ask, solve, generate: Self-evolving unified multimodal understanding and generation via self-consistency rewards.arXiv preprint arXiv:260...

    2026 arXiv

  20. [28]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms.Advances in Neural Information Processing Systems, 37:87310–87356, 2024

    Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai C Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms.Advances in Neural Information Processing Systems, 37:87310–...

    2024

  21. [29]

    Paying more attention to visual tokens in self-evolving large multimodal models

    Shravan Venkatraman, Ritesh Thawkar, Omkar Thawakar, Rao Muhammad Anwer, Hisham Cholakkal, Salman Khan, and Fahad Khan. Paying more attention to visual tokens in self-evolving large multimodal models. In ECCV. Springer, 2026

    2026

  22. [30]

    Vision-Zero: Scalable VLM self-improvement via strategic gamified self-play, 2025

    Qinsi Wang, Bo Liu, Tianyi Zhou, Jing Shi, Yueqian Lin, Yiran Chen, Hai Helen Li, Kun Wan, and Wentian Zhao. Vision-Zero: Scalable VLM self-improvement via strategic gamified self-play, 2025. arXiv preprint arXiv:2509.25541

    2025

  23. [31]

    When models judge themselves: Unsupervised self-evolution for multimodal reasoning, 2026

    Zhengxian Wu, Kai Shi, Chuanrui Zhang, et al. When models judge themselves: Unsupervised self-evolution for multimodal reasoning, 2026. arXiv preprint arXiv:2603.21289

    2026

  24. [32]

    Realworldqa

    xAI and visheratin. Realworldqa. https://huggingface.co/datasets/visheratin/realworldqa, 2024

    2024

  25. [33]

    RISE: Reliable improvement in self-evolving vision-language models, 2026

    Chaoran Xu, Yingmao Miao, Pengfei Zhang, Hao Dou, Lei Sun, and Xiangxiang Chu. RISE: Reliable improvement in self-evolving vision-language models, 2026. arXiv preprint arXiv:2605.20914

    2026 arXiv

  26. [34]

    OGLS-SD: On-policy self-distillation with outcome-guided logit steering for LLM reasoning, 2026

    Yuxiao Yang, Xiaoyun Wang, and Weitong Zhang. OGLS-SD: On-policy self-distillation with outcome-guided logit steering for LLM reasoning, 2026. arXiv preprint arXiv:2605.12400

    2026 arXiv

  27. [35]

    R1-ShareVL: Incentivizing reasoning capability of multimodal large language models via Share-GRPO, 2025

    Huanjin Yao, Qixiang Yin, Jingyi Zhang, Min Yang, Yibo Wang, Wenhao Wu, Fei Su, Li Shen, Minghui Qiu, Dacheng Tao, and Jiaxing Huang. R1-ShareVL: Incentivizing reasoning capability of multimodal large language models via Share-GRPO, 2025. arXiv preprint arXiv:2505.16673

    2025 arXiv

  28. [36]

    DAPO: An open-source LLM reinforcement learning system at scale, 2025

    Qiying Yu, Zheng Zhang, Ruofei Zhu, et al. DAPO: An open-source LLM reinforcement learning system at scale, 2025. arXiv preprint arXiv:2503.14476

    2025 arXiv

  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. arXiv preprint arXiv:2605.18740

    2026 arXiv

  30. [38]

    Self-rewarding language models

    Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, and Jason Weston. Self-rewarding language models. InInternational Conference on Machine Learning, 2024

    2024

  31. [39]

    Eric Zelikman, Yuhuai Wu, Jesse Mu, and Noah D. Goodman. STaR: Bootstrapping reasoning with reasoning. InAdvances in Neural Information Processing Systems, 2022

    2022

  32. [40]

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

    Jingyi Zhang, Jiaxing Huang, Huanjin Yao, Shunyu Liu, Xikun Zhang, Shijian Lu, and Dacheng Tao. R1-VL: Learning to reason with multimodal large language models via step-wise group relative policy optimization,

  33. [41]

    Lmms-eval: Reality check on the evaluation of large multimodal models, 2024

    Kaichen Zhang, Bo Li, Peiyuan Zhang, Fanyi Pu, Joshua Adrian Cahyono, Kairui Hu, Shuai Liu, Yuanhan Zhang, Jingkang Yang, Chunyuan Li, and Ziwei Liu. Lmms-eval: Reality check on the evaluation of large multimodal models, 2024

    2024

  34. [42]

    Self-distilled reasoner: On-policy self-distillation for large language models, 2026

    Siyan Zhao, Zhihui Xie, Mengchen Liu, Jing Huang, Guan Pang, Feiyu Chen, and Aditya Grover. Self-distilled reasoner: On-policy self-distillation for large language models, 2026. arXiv preprint arXiv:2601.18734

    2026 arXiv

  35. [43]

    ROSD: Reflective on-policy self-distillation for language model reasoning across domains, 2026

    Ziqi Zhao, Xinyu Ma, Liu Yang, Yujie Feng, Daiting Shi, Jingzhou He, Xin Xin, Zhaochun Ren, and Xiao- Ming Wu. ROSD: Reflective on-policy self-distillation for language model reasoning across domains, 2026. arXiv preprint arXiv:2605.28014. 13

    2026 arXiv

  36. [2024]

    arXiv preprint arXiv:2403.20330

    arXiv

  37. [2025]

    arXiv preprint arXiv:2503.12937

    arXiv

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.