List of Papers Browse by Subject Areas Author List
Abstract
Recent advancements in enhancing the medical reasoning capabilities of Vision-Language Models (VLMs) have predominantly relied on rule-based reinforcement learning (RL) paradigms. However, such approaches are often constrained by their dependence on \textit{sparse answer-level rewards}, which fail to provide dense supervisory signals for intermediate reasoning steps, and by the \textit{computational inefficiency inherent} in requiring extensive sampling for policy exploration.
To address these limitations, we propose \textit{On-Policy Chain-of-Visual-Thought Distillation} (OPVD) for vision-language medical reasoning, a concise yet efficient framework designed to imbue VLMs with robust medical reasoning abilities. OPVD introduces an interleaved vision-language Chain-of-Visual-Thought (CoVT) mechanism to construct reliable and granular reasoning trajectories. Unlike previous approaches, our OPVD leverages these trajectories to perform token-level dense supervision, effectively distilling multi-modal CoVT knowledge into the model via an on-policy self-distillation strategy where the teacher and student share the same policy model. This eliminates the need for external teachers or costly sampling procedures while maximizing data efficiency.
Extensive experiments conducted on the public multi-modal medical reasoning dataset demonstrate that OPVD remarkably outperforms existing models trained via standard supervised fine-tuning and RL methods, establishing a new state-of-the-art in efficient and accurate medical visual reasoning.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3650_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ZhaNin_OPVD_MICCAI2026,
author = { Zhang, Ningyue AND Yang, Qiushi},
title = { { OPVD: On-Policy Chain-of-Visual-Thought Distillation for Medical Reasoning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16878},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This manuscript proposes OPVD, efficient self-distillation framework for medical VLMs that utilizes shared-parameter teacher/student branches to provide token-level dense supervision on interleaved vision-language reasoning trajectories.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
This manuscript tackles key limitations of RL-based medical reasoning and proposes innovative CoVT mechanism. Strong empirical results.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Evaluation relies on single public dataset limiting generalizability claims. No compubation comparison (e. g. , FLOPs/samples versus RL baselines) No analysis of teacher quality or failure cases.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
methodological novelty.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
This paper introduces On-Policy Visual CoT Distillation (OPVD), a framework for training vision-language models to perform medical reasoning. The framework uses an oracle teacher to generate reasoning traces, which are used to optimize the student model’s parameters via KL divergence. The method eliminates multi-sampling that is usually required for RL-based methods. The authors demonstrate the effectiveness of their method on various medical visual reasoning tasks.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Since the teacher model is initialized from the same VLM, the oracle reasoning traces are closer to the student model’s distribution, enabling more guided learning. 2.The method outperforms bigger zero-shot models as well as other models trained using RL, setting a new state-of-the-art. 3.The paper is generally well written and easy to follow.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Missing comparison with RL-based approaches: The paper states that OPVD significantly outperforms SFT and RL methods. Including comparisons with RL-optimized Qwen-VL models would help support this claim. 2.Framing and discussion: The claim that OPVD significantly outperforms RL methods is not clearly supported by the reported results (1–2% improvement). The comparison could be presented more conservatively. 3.Currently, the paper does not provide in-depth performance and failure analysis. It would be useful to include a) cases where OPVD outperforms SFT and b) where visual CoT outperforms text-only CoT.
- Please rate the clarity and organization of this paper
Satisfactory
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
1.(Clarification) Based on my understanding, the sampling parameter in Table 1 refers to training sampling. Please clarify this in the paper.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall, I liked the paper and recommend acceptance if the authors address the weaknesses and include comparison with RL-based finetuning and in-depth analysis.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
My concerns have been clarified in the rebuttal. I remain supportive of the paper.
Review #3
- Please describe the contribution of the paper
This paper proposes OPVD, an on-policy self-distillation framework for medical vision-language reasoning. The key idea is to replace sparse reward-based RL supervision with token-level dense supervision derived from Chain-of-Visual-Thought trajectories. The method uses a teacher branch with privileged answer information and an iterative crop-and-zoom mechanism to construct higher-quality reasoning paths, which are then distilled into a student branch under standard inference conditions. The paper shows that this approach can improve both reasoning quality and training efficiency compared with standard supervised fine-tuning and recent RL-based medical reasoning methods
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
First, the paper addresses a relevant and timely problem in medical VLM reasoning, namely the inefficiency of RL-based training caused by sparse rewards and repeated sampling. The proposed use of token-level dense supervision through self-distillation is a meaningful conceptual contribution. Second, the method is clearly motivated and reasonably designed: the combination of privileged-answer teacher guidance, interleaved visual-text reasoning, and crop-and-zoom refinement is intuitive and well aligned with medical image understanding. Third, the empirical results are strong for the presented setup. OPVD consistently outperforms SFT baselines and also surpasses recent RL-based methods while requiring fewer sampling steps, which suggests a favorable efficiency-performance tradeoff. Finally, the paper includes ablations on the visual reasoning component and the answer-weighting term, which helps support the importance of the proposed design choices.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
The main weakness is that the experimental validation is still relatively limited. The model is trained on only 2,000 MRI samples and evaluated on 900 samples across MRI, CT, and X-ray, which may not be sufficient to establish broad generalizability for medical reasoning. In addition, the comparison setting may not be fully controlled, since the paper states that the exact training splits of MedVLM-R1 were unavailable and the authors constructed their own subsets, making fairness of comparison somewhat uncertain. Another concern is that the teacher branch uses privileged ground-truth answers during trajectory construction, which is effective for training but also raises questions about how much of the improvement comes from richer supervision rather than improved reasoning alone. The paper would also benefit from more extensive analysis, such as additional benchmarks, more diverse modalities/tasks, stronger ablations, and evaluation beyond multiple-choice accuracy.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The authors claimed to release the source code and/or dataset upon acceptance of the submission.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
My overall assessment is based on a positive view of the paper’s core idea and empirical performance, tempered by some concerns about evaluation breadth. I think the proposed OPVD framework is interesting because it offers a practical alternative to RL-based medical reasoning by using dense token-level supervision and an intuitive visual refinement mechanism.
- Reviewer confidence
Somewhat confident (2)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
Rebuttal
R1Q1 & R3Q1: Generalizability. We extended evaluation to the large-scale PMC-VQA dataset, including 227k samples, using Qwen2-VL-2B backbone. OPVD achieves a score of 48.7, consistently outperforming other competitors like Med-R1-2B (47.4) and MedVLM-R1-2B (47.5). Crucially, our main experiments train exclusively on MRI data while testing on CT and X-ray, demonstrating robust OOD generalization. This cross-modality and cross-benchmark consistency confirms that OPVD learns transferable, domain-agnostic reasoning patterns rather than overfitting to dataset-specific biases.
R1Q2 & R2Q4: Computation and sampling num. Toward Qwen2-VL-2B as the base model, OPVD totally requires only ~0.5 EFLOPs during training, which is approximately 6× lower than the ~3.3 EFLOPs consumed by rule-based RL baselines. The ‘‘sampling num.’’ in Table 1 denotes forward times per training iteration. OPVD requires only 2 forward passes, whereas RL baselines need 4-16 rollouts.
AC-Q1 & AC-Q2 & R1Q3 & R2Q3: Teacher quality and failure analysis. Teacher trajectories are validated via clinical priors, and Fig. 2 show the cases of the reliable reasoning process of OPVD. Failures mainly occur in complex OOD X-rays where pathologies are obscured, highlighting a need for future omni-modality training recipe.
AC-Q1 & R2Q1: RL-based Qwen comparison. To ensure a strictly fair comparison, we fine-tuned Qwen2-VL-2B with GRPO under identical training budgets, data splits, learning rates, and batch configurations. Despite matched computational resources, the GRPO baseline achieves an average accuracy of 76.41, falling significantly short of OPVD’s 80.96.This gap stems from fundamental optimization differences: GRPO relies on sparse, answer-level rewards that provide weak gradient signals for intermediate steps, often leading to reward hacking. In contrast, OPVD’s token-level distillation directly aligns the student’s generative distribution with a clinically validated teacher trajectory at each decoding step, enabling precise, dense correction that is substantially more data-efficient and stable for medical reasoning.
AC-Q1 & R2Q2: Conservative framing of gains. We will revise the claims to conservatively frame the 1-2% improvement in the manuscript. To rigorously validate reliability, we conducted paired statistical tests across multiple random seeds, yielding a p-value of 0.004, smaller than 0.01, when comparing OPVD against the strongest baseline (ViTAR), confirming the improvement is statistically significant rather than due to random variance.
AC-Q1 & R3Q2: Fair comparison. To guarantee comparison fairness, we re-implemented and re-evaluated all cited baselines using a standardized protocol: identical train/validation/test splits, consistent image preprocessing, unified hyperparameter search spaces, and the same evaluation metrics and random seeds. Under these strictly controlled conditions, OPVD consistently maintains a performance advantage, reinforcing the reproducibility of our findings.
AC-Q1 & R3Q3: Analysis on privileged prior. To isolate the impact of privileged priors, upon Qwen2-VL-2B, we trained standard SFT and RL baselines directly on the high-quality teacher-generated CoVT trajectories (i.e., using the same privileged reasoning paths as OPVD’s distillation targets). Despite access to identical prior knowledge, both baselines underperform significantly, achieving 64.29 (SFT) and 73.83 (RL) versus OPVD’s 80.96.This clear gap demonstrates that OPVD’s advantage does not stem merely from richer supervision, but from its dynamic on-policy distillation mechanism.
R3Q4: Beyond MC accuracy and benchmarks. We plan to extend our research on report generation and more diverse benchmarks in future work. The interleaved CoVT architecture is inherently well-suited for these extensions, as it already produces structured, interpretable reasoning chains that can be seamlessly adapted to free-text generation.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Although the three reviewers are rather positive about this paper, they raised some concerns about the experiment and evaluation design. As the meta-reviewer, I would like to see their responses to the following concerns: 1.Some of the reviewers have concerns about exps and results: the fairness of the comparison (different training split), the marginal improvement (1-2%), failure analysis, the ablation of richer supervision vas reasoning alone etc. 2.The evaluation pipeline seems to borrow a lot from the paper MedVLM-R1.It is acceptable for MedVLM-R1 to only evaluate on the final answer accuracy since it is one of the first works to carry out VLM reasoning. But since the field of reasoning has advanced significantly and the paper’s highlight is its CoVT, it is now also necessary to verify the intermediate process, i.e., whether the reasoning chain is plausible.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Reject
- Please justify your recommendation.
This paper has a clear motivation, while concerns of the comparison fairness and the evaluation pipeline are not well addressed, and the performance is still a bit too marginal. This paper is on the borderline.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All reviewers were satistfied with the rebuttal. The authors should address the remaining concerns in the camera-ready version.
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The reviewers are consistently supportive of the paper. OPVD offers a practical alternative to RL-based medical reasoning through on-policy visual-thought distillation, and the empirical results are viewed positively. Some concerns about evaluation breadth and novelty remain, but they do not undermine the main contribution.