Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

A counterfactual measure of evidence dependence lets on-policy distillation concentrate on diagnosis-critical tokens, improving medical visual question answering accuracy by 5.24 points over supervised fine-tuning on the paper's test subset

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

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

T0 review · deepseek-v4-flash

2026-08-02 06:31 UTC pith:YYKC2OS3

load-bearing objection Useful incremental OPD recipe for Med-VLMs, but the 'evidence-aware' mechanism is unvalidated and the evaluation needs error bars. the 3 major comments →

arxiv 2607.16303 v1 pith:YYKC2OS3 submitted 2026-07-14 cs.CV cs.AI

Med-OPD: Improving Medical Vision-Language Models via Evidence-Aware On-Policy Distillation

Yunhang Qian , Jiaquan Yu , Jiawei Liu , Meng Wang , Hongwei Bran Li , Xiaobin Hu This is my paper
classification cs.CV cs.AI
keywords medical vision-language modelon-policy distillationevidence-aware supervisioncounterfactual signalvisual groundingmedical visual question answeringtoken-level reweightingKL 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.

The paper argues that medical vision-language models often answer plausibly while ignoring the sparse, localized visual evidence that actually determines a diagnosis, and that standard on-policy distillation worsens this by spreading training signal evenly across long clinical narratives. It introduces Medical Evidence Advantage (MEA), a token-level counterfactual score that measures how much a token's likelihood drops when the image's fine detail is degraded. Med-OPD, the proposed framework, uses MEA to reweight distillation at both token and trajectory levels, focusing the teacher's supervision on diagnosis-critical tokens and evidence-reliant rollouts. On four OmniMedVQA subsets, the paper reports consistent gains over both SFT and standard OPD, with average accuracy improving by 5.24 and 2.83 percentage points respectively. If correct, this shows that where distillation is concentrated matters as much as the amount of supervision.

Core claim

The paper claims that medical evidence in VLM generations is highly concentrated: in student rollouts, most tokens have near-zero MEA and function as linguistic scaffolding, while a small top fraction carries nearly all evidence-dependent signal. Standard on-policy distillation averages KL divergence uniformly over tokens, so these sparse diagnosis-critical tokens are diluted by abundant clinical narrative. Med-OPD instead computes MEA for every token by comparing teacher log-likelihood under the original image and an evidence-degraded version (downsampled to 10% resolution then bilinearly upsampled), and reallocates the distillation objective: tokens are grouped into high- and low-evidence

What carries the argument

The load-bearing object is Medical Evidence Advantage (MEA), defined as a_t = max(log π_T(ŷ_t | v, q', ŷ_<t) − log π_T(ŷ_t | v', q', ŷ_<t), 0), where v is the original image, v' is the evidence-degraded image, and q' is the question augmented with the correct answer as a hint. This score is a teacher-grounded counterfactual: if a token becomes much more likely when fine detail is present, it is judged evidence-dependent. MEA drives two reweighting mechanisms: a token-level grouped KL loss that isolates the top ρ fraction of high-MEA tokens, and a trajectory-level softmax weight over normalized average-MEA per rollout, so that samples genuinely requiring fine-grained visual reading receive mo

Load-bearing premise

The claim rests on the assumption that the evidence-degradation operator (downsampling to 10% resolution and bilinear upsampling) selectively removes fine-grained diagnostic detail while preserving global anatomy and modality, so that the teacher's likelihood drop isolates genuine medical evidence dependence rather than generic sensitivity to image degradation.

What would settle it

A reader could take a set of student rollouts with expert-annotated lesion regions and compare the MEA ranking of tokens against those annotations: if high-MEA tokens frequently fall outside the clinically relevant structures, or if replacing the 10%-downsample corruption with a targeted lesion-masking corruption substantially reshuffles the rankings, then MEA is not measuring diagnosis-critical evidence and the reported gains would need another explanation.

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

If this is right

  • If MEA correctly identifies diagnosis-critical tokens, medical VLMs trained with Med-OPD should rely more on actual lesion features and less on language priors or clinical templates, reducing visually unsupported hallucinations.
  • The method transfers teacher capability through output distributions on student-generated rollouts, so it can improve a student model without redistributing raw patient images or expert reasoning traces, preserving privacy compatibility.
  • The reported gains on CT, MRI, Disease Diagnosis, and Lesion Grading suggest the approach generalizes across modalities and clinical question types, not just one benchmark niche.
  • Evidence-aware post-training appears complementary to model scaling: the 2B student with Med-OPD improves over its SFT baseline and approaches or exceeds some larger zero-shot models, indicating supervision allocation matters independently of parameter count.
  • The token-level MEA scoring can be applied on top of any on-policy distillation setup, making it a drop-in reweighting scheme for future student-teacher pipelines.

Where Pith is reading between the lines

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

  • MEA itself could be repurposed as an evaluation or interpretability tool: after training, the same counterfactual comparison could highlight which tokens a model depends on, giving clinicians a per-token evidence-attribution map without additional annotations.
  • The evidence-degradation operator (10% downsample plus bilinear upsampling) is an unvalidated proxy for removing diagnostic evidence; replacing it with lesion-region masking or clinical annotations would test whether MEA tracks semantic evidence or just low-level texture sensitivity.
  • Because MEA is dense and token-level, it could serve as a shaped reward in reinforcement-learning pipelines, connecting evidence reliance to outcome optimization more tightly than scalar answer rewards do.
  • The hyperparameters ρ and λ, fixed in the paper, may need per-modality tuning; the long-tail distribution of MEA likely varies with imaging modality and question type, so a more adaptive grouping rule could further improve the method.

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 Med-OPD, a post-training method for medical vision-language models that augments on-policy distillation with a Medical Evidence Advantage (MEA) signal. MEA is computed for each token by comparing the teacher's log-likelihood under the original image and an evidence-degraded version (downsampled to 10% then bilinearly upsampled), and is used to reweight the distillation loss at both trajectory and token levels. Experiments on four OmniMedVQA subsets with a Qwen3-VL-2B student and Qwen3-VL-4B teacher report average gains of 5.24 percentage points over SFT and 2.83 percentage points over standard OPD. The loss formulation is coherent and code/data are released, but the evidence-aware interpretation rests on an unvalidated degradation operator, and the empirical support is statistically thin.

Significance. If the central claim holds, Med-OPD would provide a practical way to concentrate distillation supervision on diagnosis-critical tokens, potentially improving medical VLM grounding without requiring evidence annotations. The paper has clear strengths: the method is precisely specified in Eqs. (4)-(10), evaluation is on an external benchmark (OmniMedVQA, not a constructed test set), the code and data are public, and the use of on-policy distillation is well-motivated for privacy-sensitive medical settings. However, the core novelty is MEA, and that signal is defined through an image degradation operator whose validity is not demonstrated. The current evidence does not establish that the gains come from 'evidence awareness' rather than from a generic reweighting heuristic. The paper is worth serious consideration, but the central empirical claim needs substantially more validation.

major comments (3)
  1. [§3.2, Eq. (4)] The MEA signal is defined by the teacher log-likelihood gap between the original image and v', where v' is created by downsampling to 10% resolution and bilinear upsampling. This operator attenuates all high-frequency content globally—edges, contrast, and texture—not only diagnostic evidence. The claim that it 'largely preserves global anatomy and modality' is not supported, so Eq. (4) may measure generic blur sensitivity rather than dependence on diagnosis-critical evidence. This is load-bearing because the paper's central claim is that Med-OPD is evidence-aware. Please validate with (a) ground-truth evidence masks or expert annotations, (b) an ablation over degradation operators (e.g., Gaussian blur, ROI masking, different downsampling ratios), and (c) a control showing MEA-based grouping outperforms random token grouping with matched group sizes. Figure 3 is anecdotal and insufficient
  2. [Table 1, §4.2] Each subset contains only 300 test questions, and all results come from a single training run with no error bars or significance tests. For an accuracy around 0.7, the binomial standard error is about ±2.6 pp, so the per-subset gains over standard OPD (1.3–5.0 pp) are within plausible noise. The headline average gain of 2.83 pp over OPD is also not shown to be significant. Please report multiple seeds with means and variances, or at least bootstrap confidence intervals, and state whether the average improvement is statistically reliable.
  3. [§3.3–§3.5, §4.1] The method introduces hyperparameters ρ, λ, τ, K, and degradation_ratio, but §4.1 reports none of their values, and there is no ablation isolating the token-grouping term (Eq. 7) from the trajectory weighting (Eqs. 8–9) or showing sensitivity to these choices. Without a control that replaces MEA-based groups with random groups or a non-evidence heuristic (e.g., upweighting rare content words), the accuracy gains could be produced by any token-grouping scheme. This is a specific, fixable gap, but it is central to the novelty claim.
minor comments (5)
  1. [§3.4, Eq. (8)] The small constant ε is mentioned but never defined. Please specify its value or state that it is a fixed small constant.
  2. [§4.1] 'Each training run lasts for up to 10 epochs, with a maximum of 100 or 200 training steps depending on the experimental setting' is vague. Specify which setting uses 100 versus 200 steps.
  3. [Figure 5] The generated example outputs contain numerous typographical artifacts (e.g., 'fetniled', 'tisstues', 'jounts', 'ansver', 'sane level', 'sofi tissue gontrast'). If these are verbatim model outputs, this should be stated; the current presentation looks like OCR errors and undermines readability.
  4. [§4.1] 'Student Top-K distillation with ktop = 16' is referenced but not described. Please explain what Top-K distillation means here, or move the reference to the first occurrence.
  5. [Abstract] Minor grammar: 'The source code and data is publicly available' should be 'are publicly available.'

Circularity Check

0 steps flagged

No significant circularity: the central accuracy claim is benchmarked externally, and MEA's degradation assumption is a validity gap, not a tautology.

full rationale

Med-OPD's headline result—improving average accuracy from 0.6775 (standard OPD) and 0.6534 (SFT) to 0.7058 (Med-OPD) on OmniMedVQA subsets (Table 1, §4.2)—is evaluated on held-out test examples from an external benchmark. The final objective (Eq. 10) reweights the standard OPD divergence by MEA-derived weights; it does not fit any parameter to the test set and then report that fit as a prediction. The MEA definition (Eq. 4) is an operational counterfactual: it measures teacher likelihood differences between original and downsampled images. That measure is not defined in terms of the accuracy metric, and the training signal produced from it cannot by construction force the reported accuracy gains. The paper's weakest point is that the evidence-degradation operator (downsampling to 10% + bilinear upsampling) is asserted, not validated, to selectively remove diagnosis-critical visual evidence; this is a real confounding/validity concern, and the absence of degradation ablations or ground-truth evidence masks means the 'evidence-aware' interpretation is under-supported. However, an unverified assumption is not the same as circularity: Eq. (4) does not presuppose that the likelihood gap corresponds to clinical relevance, and the paper does not use MEA both as the definition of 'evidence' and as the sole evidence that it finds 'evidence' in a way that reduces to a fit. Self-citations ([10], [27], [35]) appear in related-work or survey contexts and are not load-bearing for the derivation. Therefore, no step in the presented derivation chain reduces to its own inputs by construction.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The method rests on two hand-defined measurement choices (degradation operator and clipped likelihood difference) plus three undisclosed weighting hyperparameters. None of these is fitted to OmniMedVQA in an obvious way, so the central accuracy result has external grounding; the evidence-groundedness claim itself is not independently verified.

free parameters (5)
  • rho (ρ)
    Fraction of tokens assigned to the high-MEA group (Eq. 5); controls how aggressively evidence tokens are separated; never given a value in the paper.
  • lambda (λ)
    Balancing coefficient between high- and low-MEA group KL terms (Eq. 7); value undisclosed.
  • tau (τ)
    Softmax temperature for trajectory-level weights (Eq. 9); value undisclosed.
  • degradation_ratio = 10%
    Spatial downsampling ratio used to construct the evidence-degraded image v' (§3.2); hand-chosen, no ablation reported.
  • K = 8
    Number of on-policy rollouts per training example (§4.1); affects the trajectory weighting statistics.
axioms (5)
  • standard math Reverse-KL distillation on student-generated prefixes is a valid, stable supervision signal for transferring teacher behavior (standard OPD, cited [1,7]).
    Invoked in Eq. (3) as the foundation of the method; not proved here, taken from prior work.
  • domain assumption The clipped teacher log-likelihood difference in Eq. (4) measures per-token dependence on medical visual evidence.
    Core identifying assumption of MEA; not validated against any ground-truth evidence annotation.
  • domain assumption Downsampling to 10% then bilinear upsampling selectively destroys fine-grained diagnostic evidence while preserving global anatomy and modality.
    §3.2; if false, MEA confounds evidence dependence with other image statistics; no ablation supports it.
  • ad hoc to paper Clipping the evidence difference at zero discards no meaningful signal (Eq. 4).
    max(·,0) is a design choice; negative differences (tokens more likely under degraded image) are treated as zero evidence dependence without justification.
  • domain assumption Student Top-K distillation (k_top=16) from [18] is an appropriate distillation variant.
    Adopted from the cited recipe (§4.1) without ablation in the medical setting.
invented entities (1)
  • Medical Evidence Advantage (MEA) no independent evidence
    purpose: Token-level scalar quantifying each token's dependence on fine-grained medical visual evidence; drives both token grouping (Eq. 5) and trajectory weighting (Eqs. 8–9).
    New quantity defined in Eq. (4). The paper provides qualitative examples but no external grounding benchmark or falsifiable prediction that high-MEA tokens correspond to annotated clinical evidence; independent_evidence is therefore false.

pith-pipeline@v1.3.0-alltime-deepseek · 180 in / 17813 out tokens · 202053 ms · 2026-08-02T06:31:13.754254+00:00 · methodology

0 comments
read the original abstract

Medical Vision-Language Models (Med-VLMs) require reliable reasoning from fine-grained visual evidence, yet existing models can produce plausible clinical answers by relying on language priors or medical templates rather than truly attending to diagnosis-critical regions. On-Policy Distillation (OPD) offers dense token-level supervision on student-generated trajectories and provides a privacy-compatible means of capability transfer without requiring the redistribution of raw patient data. However, standard OPD uniformly distills all tokens, causing sparse evidence-dependent tokens to be diluted by abundant clinical narrative tokens. Inspired by the success of OPD in the large language model community, we propose \textbf{Med-OPD}, to our knowledge the first unified post-training framework that integrates on-policy distillation with medical evidence-aware supervision for Med-VLMs. We introduce \textbf{Medical Evidence Advantage} (MEA), a teacher-grounded counterfactual signal that uses an answer-aware hint to focus teacher scoring on evidence supporting the target diagnosis, and measures each token's dependence on medical visual evidence by comparing teacher likelihoods under the original and evidence-degraded imaging modalities. Based on MEA, Med-OPD redistributes the distillation signal at both the token and trajectory levels, emphasizing diagnosis-critical tokens and evidence-reliant rollouts. Experiments on OmniMedVQA subsets show that Med-OPD consistently outperforms SFT and standard OPD across CT, MRI, Disease Diagnosis, and Lesion Grading. These results demonstrate that evidence-aware distillation can better strengthen medical VLMs' reliance on key visual evidence and improve reliable multimodal medical reasoning. The source code and data is publicly available at: https://github.com/yunhang8658/MedOPD.git

Figures

Figures reproduced from arXiv: 2607.16303 by Hongwei Bran Li, Jiaquan Yu, Jiawei Liu, Meng Wang, Xiaobin Hu, Yunhang Qian.

Figure 1
Figure 1. Figure 1: Common paradigms for improving medical VLM reasoning. SFT learns from labeled examples, CoT elicits intermediate reasoning steps, outcome-reward RL optimizes scalar feedback, and OPD transfers teacher behavior to a student policy through distillation. than redistributing sensitive patient data. However, standard OPD uniformly applies KL distillation to all tokens, while many tokens in medical responses are… view at source ↗
Figure 2
Figure 2. Figure 2: MEA is long-tailed. Left: the top 10% high-MEA tokens carry most of the total MEA mass. Right: most tokens have near-zero MEA, while only a small subset strongly depends on medical visual evidence. reliance on key medical visual evidence and validating its effectiveness on medical multimodal question answering tasks. 2 Related Work 2.1 Medical Vision-Language Models and Medical Reasoning Med-VLMs have rapi… view at source ↗
Figure 3
Figure 3. Figure 3: High-MEA tokens align with fine-grained medical evidence. High-MEA tokens concentrate on clinically relevant details and diagnosis-critical findings, while low-MEA tokens mainly correspond to general narrative. Given a training triple (v, q, y) ∼ S, where v is a medical image, q is a question, and y = (y1, . . . , y|y|) is the target answer, supervised fine-tuning (SFT) trains the student by teacher forcin… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of Med-OPD. Med-OPD first constructs an evidence-degraded view of the medical image and an answer-aware teacher prompt. The student generates on-policy rollouts from the original image-question pair, while the teacher evaluates these rollouts under both full-evidence and degraded-evidence conditions to estimate MEA. The resulting MEA signal is used to reallocate distillation strength across rollou… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of evidence-grounded reasoning. Token-level highlights on a representative MRI modality question. Med-OPD more strongly emphasizes image-grounded descriptors, including internal body structures and soft-tissue contrast, than the baseline and standard OPD. the student samples K = 8 candidate answers for each training example, which are used to compute MEA-guided trajectory-level weigh… view at source ↗

discussion (0)

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

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

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

    cs.CV 2026-07 conditional novelty 6.0

    Counterfactual present/removed teacher views attribute visually supported corrections and reconstruct student-anchored distillation targets that beat source-mixed multimodal OPD.

Reference graph

Works this paper leans on

37 extracted references · 13 linked inside Pith · cited by 1 Pith paper

  1. [1]

    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. InInterna- tional Conference on Learning Representations, volume 2024, pages 21246–21263, 2024

    2024

  2. [2]

    Towards injecting medical visual knowledge into multimodal llms at scale

    Junying Chen, Chi Gui, Ruyi Ouyang, Anningzhe Gao, Shunian Chen, Guiming Hardy Chen, Xidong Wang, Zhenyang Cai, Ke Ji, Xiang Wan, et al. Towards injecting medical visual knowledge into multimodal llms at scale. InProceedings of the 2024 conference on empirical methods in natural language processing, pages 7346–7370, 2024

    2024

  3. [3]

    Mmedexpert-r1: Strengthening multimodal medical reasoning via domain-specific adaptation and clinical guide- line reinforcement.arXiv preprint arXiv:2601.10949, 2026

    Meidan Ding, Jipeng Zhang, Wenxuan Wang, Haiqin Zhong, Xiaoling Luo, Wenting Chen, and Linlin Shen. Mmedexpert-r1: Strengthening multimodal medical reasoning via domain-specific adaptation and clinical guide- line reinforcement.arXiv preprint arXiv:2601.10949, 2026

    arXiv 2026

  4. [4]

    Multi-modal hallucination control by visual information grounding

    Alessandro Favero, Luca Zancato, Matthew Trager, Siddharth Choudhary, Pramuditha Perera, Alessandro Achille, Ashwin Swaminathan, and Stefano Soatto. Multi-modal hallucination control by visual information grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14303–14312, 2024

    2024

  5. [5]

    Visual description grounding reduces hallucinations and boosts reasoning in lvlms

    Sreyan Ghosh, Chandra Kiran Evuru, Sonal Kumar, Utkarsh Tyagi, Oriol Nieto, Zeyu Jin, and Dinesh Manocha. Visual description grounding reduces hallucinations and boosts reasoning in lvlms. InInternational Conference on Learning Representations, volume 2025, pages 66510–66547, 2025

    2025

  6. [6]

    Med-cmr: A fine-grained benchmark integrating visual evidence and clinical logic for medical complex multimodal reasoning

    Haozhen Gong, Xiaozhong Ji, Yuansen Liu, Wenbin Wu, Xiaoxiao Yan, Jingjing Liu, Kai Wu, Jiazhen Pan, Bailiang Jian, Jiangning Zhang, et al. Med-cmr: A fine-grained benchmark integrating visual evidence and clinical logic for medical complex multimodal reasoning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 412...

    2026

  7. [7]

    Minillm: Knowledge distillation of large language models

    Yuxian Gu, Li Dong, Furu Wei, and Minlie Huang. Minillm: Knowledge distillation of large language models. In International Conference on Learning Representations, volume 2024, pages 32694–32717, 2024

    2024

  8. [8]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, Qihao Zhu, Runxin Xu, Ruoyu Zhang, Shirong Ma, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    Uav-marl: Multi-agent reinforcement learning for time-critical and dynamic medical supply delivery.arXiv preprint arXiv:2603.10528, 2026

    Islam Guven and Mehmet Parlak. Uav-marl: Multi-agent reinforcement learning for time-critical and dynamic medical supply delivery.arXiv preprint arXiv:2603.10528, 2026

    arXiv 2026

  10. [10]

    The landscape of medical agents: A survey

    Xiaobin Hu, Yunhang Qian, Jiaquan Yu, Jingjing Liu, Xiaozhong Ji, Chengming Xu, Peng Tang, Chengming Xu, Peng Tang, Jiawei Liu, et al. The landscape of medical agents: A survey. 2026

    2026

  11. [11]

    Omnimedvqa: A new large-scale comprehensive evaluation benchmark for medical lvlm

    Yutao Hu, Tianbin Li, Quanfeng Lu, Wenqi Shao, Junjun He, Yu Qiao, and Ping Luo. Omnimedvqa: A new large-scale comprehensive evaluation benchmark for medical lvlm. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22170–22183, 2024

    2024

  12. [12]

    Distillm: Towards streamlined distillation for large language models.arXiv preprint arXiv:2402.03898, 2024

    Jongwoo Ko, Sungnyun Kim, Tianyi Chen, and Se-Young Yun. Distillm: Towards streamlined distillation for large language models.arXiv preprint arXiv:2402.03898, 2024

    Pith/arXiv arXiv 2024

  13. [13]

    Med-r1: Reinforcement learning for generalizable medical reasoning in vision-language models.IEEE transactions on medical imaging, 2026

    Yuxiang Lai, Jike Zhong, Ming Li, Shitian Zhao, Yuheng Li, Konstantinos Psounis, and Xiaofeng Yang. Med-r1: Reinforcement learning for generalizable medical reasoning in vision-language models.IEEE transactions on medical imaging, 2026

    2026

  14. [14]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13872–13882, 2024

    2024

  15. [15]

    Joint optimization of reasoning and dual-memory for self-learning diagnostic agent.arXiv preprint arXiv:2604.07269, 2026

    Bingxuan Li, Simo Du, and Yue Guo. Joint optimization of reasoning and dual-memory for self-learning diagnostic agent.arXiv preprint arXiv:2604.07269, 2026. 8 APREPRINT- JULY21, 2026

    Pith/arXiv arXiv 2026

  16. [16]

    Llava-med: Training a large language-and-vision assistant for biomedicine in one day

    Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36:28541–28564, 2023

    2023

  17. [17]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InInternational conference on machine learning, pages 19730–19742. PMLR, 2023

    2023

  18. [18]

    Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    Yaxuan Li, Yuxin Zuo, Bingxiang He, Jinqian Zhang, Chaojun Xiao, Cheng Qian, Tianyu Yu, Huan-ang Gao, Wenkai Yang, Zhiyuan Liu, et al. Rethinking on-policy distillation of large language models: Phenomenology, mechanism, and recipe.arXiv preprint arXiv:2604.13016, 2026

    Pith/arXiv arXiv 2026

  19. [19]

    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. InProceedings of the 2023 conference on empirical methods in natural language processing, pages 292–305, 2023

    2023

  20. [20]

    Self-rewarding vision-language model via reasoning decomposition.arXiv preprint arXiv:2508.19652, 2025

    Zongxia Li, Wenhao Yu, Chengsong Huang, Zhenwen Liang, Rui Liu, Fuxiao Liu, Jingxi Che, Dian Yu, Jordan Boyd-Graber, Haitao Mi, et al. Self-rewarding vision-language model via reasoning decomposition.arXiv preprint arXiv:2508.19652, 2025

    Pith/arXiv arXiv 2025

  21. [21]

    A survey on hallucination in large vision-language models.arXiv preprint arXiv:2402.00253, 2024

    Hanchao Liu, Wenyuan Xue, Yifei Chen, Dapeng Chen, Xiutian Zhao, Ke Wang, Liping Hou, Rongjun Li, and Wei Peng. A survey on hallucination in large vision-language models.arXiv preprint arXiv:2402.00253, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26296–26306, 2024

    2024

  23. [23]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    2023

  24. [24]

    Visual-advantage on-policy distillation for vision-language models.arXiv preprint arXiv:2605.21924, 2026

    Ruiqi Liu, Xiaolei Lv, Gengsheng Li, Ximo Zhu, Zhiheng Wang, Zhengbo Zhang, Junkai Chen, Zhiheng Li, Bo Li, Jun Gao, et al. Visual-advantage on-policy distillation for vision-language models.arXiv preprint arXiv:2605.21924, 2026

    Pith/arXiv arXiv 2026

  25. [25]

    Localizing before answering: A hallucination evaluation benchmark for grounded medical multimodal llms.arXiv preprint arXiv:2505.00744, 2025

    Dung Nguyen, Minh Khoi Ho, Huy Ta, Thanh Tam Nguyen, Qi Chen, Kumar Rav, Quy Duong Dang, Satwik Ramchandre, Son Lam Phung, Zhibin Liao, et al. Localizing before answering: A hallucination evaluation benchmark for grounded medical multimodal llms.arXiv preprint arXiv:2505.00744, 2025

    arXiv 2025

  26. [26]

    Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning

    Jiazhen Pan, Che Liu, Junde Wu, Fenglin Liu, Jiayuan Zhu, Hongwei Bran Li, Chen Chen, Cheng Ouyang, and Daniel Rueckert. Medvlm-r1: Incentivizing medical reasoning capability of vision-language models (vlms) via reinforcement learning. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 337–347. Springer, 2025

    2025

  27. [27]

    Medmaslab: A unified orchestration framework for benchmarking multimodal medical multi-agent systems.arXiv preprint arXiv:2603.09909, 2026

    Yunhang Qian, Xiaobin Hu, Jiaquan Yu, Siyang Xin, Xiaokun Chen, Jiangning Zhang, Peng-Tao Jiang, Jiawei Liu, and Hongwei Bran Li. Medmaslab: A unified orchestration framework for benchmarking multimodal medical multi-agent systems.arXiv preprint arXiv:2603.09909, 2026

    arXiv 2026

  28. [28]

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

    Stéphane 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, pages 627–635. JMLR Workshop and Conference Proceedings, 2011

    2011

  29. [29]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024.URL https://arxiv

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024.URL https://arxiv. org/abs/2402.03300, 2(3):5, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Openclaw-rl: Train any agent simply by talking.arXiv preprint arXiv:2603.10165, 2026

    Yinjie Wang, Xuyang Chen, Xiaolong Jin, Mengdi Wang, and Ling Yang. Openclaw-rl: Train any agent simply by talking.arXiv preprint arXiv:2603.10165, 2026

    Pith/arXiv arXiv 2026

  31. [31]

    Perception-aware policy optimization for multimodal reasoning (2025).URL https://arxiv

    Zhenhailong Wang, Xuehang Guo, Sofia Stoica, Haiyang Xu, Hongru Wang, Hyeonjeong Ha, Xiusi Chen, Yangyi Chen, Ming Yan, Fei Huang, et al. Perception-aware policy optimization for multimodal reasoning (2025).URL https://arxiv. org/abs/2507.06448

    Pith/arXiv arXiv 2025

  32. [32]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    2022

  33. [33]

    V-dpo: Mitigating hallucination in large vision language models via vision-guided direct preference optimization

    Yuxi Xie, Guanzhen Li, Xiao Xu, and Min-Yen Kan. V-dpo: Mitigating hallucination in large vision language models via vision-guided direct preference optimization. InFindings of the Association for Computational Linguistics: EMNLP 2024, pages 13258–13273, 2024

    2024

  34. [34]

    Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling

    Wenda Xu, Rujun Han, Zifeng Wang, Long Le, Dhruv Madeka, Lei Li, William Wang, Rishabh Agarwal, Chen-Yu Lee, and Tomas Pfister. Speculative knowledge distillation: Bridging the teacher-student gap through interleaved sampling. InInternational Conference on Learning Representations, volume 2025, pages 64616–64646, 2025. 9 APREPRINT- JULY21, 2026

    2025

  35. [35]

    Emambair: Efficient visual state space model for event-guided image reconstruction

    Wei Yu and Yunhang Qian. Emambair: Efficient visual state space model for event-guided image reconstruction. arXiv preprint arXiv:2605.08073, 2026

    Pith/arXiv arXiv 2026

  36. [36]

    Dopd: Dual on-policy distillation.arXiv preprint arXiv:2606.30626, 2026

    Xinlei Yu, Gen Li, Qingyi Si, Guibin Zhang, Yuqi Xu, Congcong Wang, Shuai Dong, Kaiwen Tuo, Xiangyu Zeng, Kaituo Feng, et al. Dopd: Dual on-policy distillation.arXiv preprint arXiv:2606.30626, 2026

    Pith/arXiv arXiv 2026

  37. [37]

    Vision-opd: Learning to see fine details for multimodal llms via on-policy self-distillation.arXiv preprint arXiv:2605.18740, 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.arXiv preprint arXiv:2605.18740, 2026. 10

    Pith/arXiv arXiv 2026