Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Large vision-language models (VLMs) show strong general reasoning capabilities but remain unreliable for medical visual question answering. Existing medical VLMs often rely on supervised fine-tuning, which is limited by data scarcity, computational cost, and the risk of parameter-update-induced forgetting. We propose EvoMed, a training-free framework that enhances medical competency without updating model parameters. EvoMed employs a closed-loop self-evolving mechanism to iteratively distill clinical priors into a compact external knowledge base via reasoning, critique, and contrastive summarization. A coarse-to-fine retrieval mechanism selects highly relevant priors to guide inference toward clinically grounded predictions. Experiments on OmniMedVQA and SLAKE show that EvoMed improves both Qwen2.5-VL-7B and Qwen3-VL-235B-A22B by up to 7.5\% using only 500 evolution samples. Code is available at \texttt{\url{https://github.com/Hunan-Tiger/EvoMed}}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/4228_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Hunan-Tiger/EvoMed

Link to the Dataset(s)

SLAKE dataset: A comprehensive semantic labels annotated by experienced physicians and a new structural medical knowledge base for Med-VQA. https://www.med-vqa.com/slake/#gt-Features OmniMedVQA dataset: A large-scale and comprehensive Visual Question Answering benchmark tailored to the medical domain. https://openxlab.org.cn/datasets/GMAI/OmniMedVQA

BibTex

@InProceedings{LiJia_EvoMed_MICCAI2026,
        author = { Li, Jiaming AND Xiong, Honglin AND Wang, Qian},
        title = { { EvoMed: Self-Evolving Medical VLM via Training-free Continued Learning } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16879},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The work proposes a training-free and cost-efficient approach that enhances interpretability by providing supporting evidence for generated outputs via retrieval-augmented generation. Additionally, the method incorporates a self-evaluation mechanism and iteratively updates the knowledge base while avoiding redundancy.

  • 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.
    • The proposed method is logically sound and well-motivated.
    • The approach operates without any parameter updates, making it computationally efficient and practical for deployment.
    • The use of publicly available datasets enhances the reproducibility and transparency of the work.
    • Both qualitative and quantitative experiments are provided.
  • 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 core components of the method, retrieval and re-ranking, depend heavily on the capability of the VLM to generate accurate visual descriptions (V_desc) and semantic triplets (T). However, the VLMs employed are general-domain models, which introduces a methodological gap. The overall performance is tightly coupled to the quality and alignment of these intermediate representations with the ground truth, yet this dependency is not sufficiently analyzed. In particular, the paper lacks a systematic evaluation of the quality and downstream impact of V_desc and T, even though all subsequent stages rely on them for filtering. Errors in these representations may therefore propagate throughout the pipeline.

    • The definition and role of the “Query Content” element within the semantic triplet T are unclear, making it difficult to fully understand the method.

    • The evaluation relies entirely on a general-domain LLM-as-a-judge, without manual validation (sanity checks such as sampling a subset of evaluated instances and verifying whether the LLM’s judgments align with human assessment). This is particularly concerning given the specialized nature of the task, where domain-specific expertise is likely required for reliable assessment.

    • There is conceptual confusion between catastrophic forgetting and accuracy. The claim that EvoMed “completely circumvents catastrophic forgetting” is not substantiated by appropriate metrics. Catastrophic forgetting is typically evaluated in a continual learning setting (e. g. , via forgetting rates across sequential tasks), whereas the paper only reports accuracy comparisons (e. g. , zero-shot vs. fine-tuned). This does not adequately support the stated claim.

    • There is no comparison with any RAG framework.

    • The random selection strategy raises concerns about coverage—specifically, whether each test sample is likely to have a relevant counterpart in the adaptation set.

    • The evaluation protocol is insufficiently explained. Although the test splits are larger (e. g. , >3,000 CT samples and >1,000 microscopy samples), the experiments are conducted on relatively small subsets (550 and 534 samples, respectively). The rationale for this selection is not provided, raising concerns about representativeness.

  • 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 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The overall score reflects a balance between a well-motivated and practical idea and several methodological concerns. The paper proposes a coherent, training-free framework with interpretability benefits and demonstrates results on public datasets.

    However, confidence in the approach is limited by its strong dependence on intermediate representations (V_desc and semantic triplets) without sufficient analysis of their quality, as well as an evaluation protocol that relies solely on a general-domain LLM-as-a-judge without human validation. Additionally, some claims (e.g., regarding catastrophic forgetting) are not properly supported, and dataset subsampling is insufficiently justified.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    Reject

  • [Post rebuttal] Please justify your final decision from above.

    The authors addressed several comments; however, some concerns still remain unresolved.

    -The claim that the larger VLM generates more faithful V_desc is not sufficiently supported, since no faithfulness metrics or qualitative examples of generated V_desc are provided.

    -If evaluation reduces to closed-set answer matching, it remains unclear why an LLM-as-a-judge is necessary instead of deterministic matching rules.

    -The revised wording “avoids parameter-update-induced forgetting risk” is more appropriate; however, avoiding parameter updates is a generic property of frozen-backbone RAG frameworks (e.g. MOTOR (MICCAI 2025)) rather than a unique characteristic of EvoMed. Forgetting rates won’t be applicable to measure since EvoMed is not a continual learning framework.



Review #2

  • Please describe the contribution of the paper

    This paper proposes EvoMed, a training-free framework for enhancing medical visual question answering (Med-VQA) without modifying any model parameters. The key idea is to replace supervised fine-tuning with a self-evolving knowledge accumulation process that builds an external knowledge base (KB) through the VLM’s own reasoning experience.

    The framework has two main components. First, a Self-Evolving Knowledge Accumulation loop operates over a small adaptation set (500 samples) in a closed-loop with four roles: Actor, Critic, Summarizer and Updater.

    Second, a Coarse-to-Fine Retrieval strategy at inference time first prunes the KB using semantic triplets and then reranks candidates using a cross-encoder, injecting the most relevant clinical priors into the VLM’s context window.

  • 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.Compelling problem formulation: The paper clearly identifies real limitations of supervised fine-tuning for medical VLMs and proposes a coherent alternative that addresses those limitations. The framing of ‘self-evolving knowledge accumulation’ as an externalized learning process that keeps model weights frozen is practically appealing, especially for clinical settings where frequent model retraining is impractical.

    2.Remarkable data efficiency: The most striking result is that EvoMed achieves substantial improvements using only 500 evolution samples and ~12-15M tokens ($10 API cost). This data efficiency is highly relevant for medical imaging where labeled data is expensive.

    3.Cross-model knowledge transferability (Table 3): This is a novel and insightful experiment. The finding that a KB built by the 7B model can improve the 72B model (and vice versa) suggests the accumulated clinical priors are genuinely model-agnostic semantic knowledge, not model-specific prompt artifacts.

    1. Thorough experimental analysis: The paper includes multiple informative analyses beyond the main results: per-modality breakdowns (Table 1), comparison with fine-tuned medical VLMs (Table 2), cross-model transfer (Table 3), evolution dynamics over epochs (Fig. 3), and a qualitative case study (Fig. 2).
  • 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.The comparison with fine-tuned medical VLMs (Table 2) is misleading: The paper claims EvoMed ‘significantly outperforms extensively fine-tuned medical VLMs,’ but this comparison is fundamentally unfair. LLaVA-Med [10] and Med-Flamingo [15] were published in 2023, built on much weaker backbone architectures (LLaMA-7B era), and trained on different data distributions.

    EvoMed uses Qwen2.5-VL (2025), which has vastly stronger base capabilities. The zero-shot Qwen2.5-VL-72B already scores 80.18% on OmniMedVQA-Rad, far exceeding LLaVA-Med’s 52.91%.

    So EvoMed’s lead against SFT Medical VLM could be primarily attributable to the stronger backbone, not the proposed method. A fairer comparison would apply EvoMed to the same backbone as competing methods, or compare against recent medical VLMs built on comparable backbones.

    1. No comparison with standard RAG baselines: EvoMed is fundamentally a retrieval-augmented generation (RAG) system with a self-evolving knowledge base. Yet the paper does not compare against simpler RAG baselines, such as static RAG using a medical textbook or using the same 500 QA pairs as knowledge source. So far, it is unclear how much of the improvement comes from the self-evolution mechanism versus simply providing any relevant medical context at inference time.
  • 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?

    On the positive side, the problem is well-motivated and the proposed solution is creative. The data efficiency result is the paper’s strongest selling point and the cross-model knowledge transfer experiment adds a novel dimension that goes beyond simply showing improved accuracy.

    On the negative side, the most concerning issue is the lack of RAG baselines, which makes it hard to determine whether the self-evolution mechanism or simply having relevant context at inference time drives the improvement. The misleading comparison with older, weaker-backbone medical VLMs weakens the paper’s claims about outperforming fine-tuning.

    In a nutshell, the conceptual contribution is awesome but the evaluation gaps are real. The rebuttal could change my assessment if the authors provide a RAG baseline ablation and acknowledge the backbone difference in the Table 2 comparison.

  • 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.

    After carefully reading the authors’ rebuttal, I maintain my positive assessment. Here is my reasoning:

    1.RAG baseline (our primary concern):

    The rebuttal provides the ablation we specifically requested. A textbook-RAG baseline using four open-access NCBI radiology textbooks with BM25 retrieval actually hurts performance (Qwen2.5-VL: 67.21% to 60.36%; Qwen3-VL: 80.18% to 77.09%), while EvoMed achieves substantial gains (+6.61 and +5.46) on identical backbones.

    This is compelling evidence that the self-evolution mechanism is the critical driver of improvement. The finding that imprecise textbook context actively distracts the model’s visual reasoning, while compact error-driven clinical priors improve it, validates the core design hypothesis.

    2.Fairness of SFT comparison (our secondary concern)

    Fairness of SFT comparison (our secondary concern): The authors concede that Table 2 is not a same-backbone comparison and commit to moderating the claim and noting backbone/training-data differences in the camera-ready. They reframe the SFT comparison as illustrating distribution overfitting: LLaVA-Med achieves 85.34% on in-distribution SLAKE but collapses to 52.91% on OmniMed-Rad. This is a fair point.

    The controlled same-backbone improvements (+6.61 for 7B, +5.46 for the larger model) remain the proper evidence for the method’s effectiveness. I find this response satisfactory.

    3.Model identification error

    The rebuttal discloses that the model reported as “Qwen2.5-VL-72B” is actually Qwen3-VL-235B-A22B, a Mixture-of-Experts model. This is a significant disclosure that changes the identity of the backbone used throughout the paper.

    While the authors argue that the activated parameters (~22B) are smaller than a dense 72B and that API costs are even lower, the fact remains that every “72B” label in Tables 1, 2, and 3 is incorrect, and the cross-model transfer results (Table 3) are between a 7B and a 235B-MoE, not a 7B and 72B as presented.

    This is more than a typo, and it requires substantial revision of the paper’s narrative.

    That said, I do not believe it invalidates the core contributions: the same-backbone improvements are still demonstrated, the 7B results (which are unaffected by this error) already show the method’s value, and the self-evolution framework design is independent of which specific backbone is used. The camera-ready must correct this thoroughly.

    In general, the RAG baseline ablation supports the paper’s thesis. The self-evolution mechanism outperforms both zero-shot inference and naive textbook retrieval, validating that the contribution is not trivially reducible to “just add RAG.”

    The model identification error is concerning and must be corrected, but it does not undermine the method’s conceptual contribution or the 7B-model results.



Review #3

  • Please describe the contribution of the paper

    This paper proposes EvoMed, a training-free framework that enhances medical competency without updating model parameters. It distills clinical priors into a compact external knowledge base and uses a coarse-to-fine retrieval mechanism to select highly relevant priors from this knowledge base, guiding inference toward more clinically grounded predictions.

  • 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.The paper explores an interesting direction for continual learning by using a training-free pipeline with an external knowledge base and retrieval mechanism, rather than relying on continual parameter tuning across downstream tasks. 2.The qualitative visualizations are clear and helpful, and they effectively illustrate the improvement brought by EvoMed’s retrieval mechanism.

  • 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.

    Unclear continued-learning setting in the experiments. The paper would benefit from clarifying several points: 1.Are CT, X-ray, and MRI in Table 1 sequential tasks or just separate adaptation subsets within OmniMed-Rad? 2.Are the reported results final accuracies after all incremental stages? 3.What sample proportion from each modality is used in the self-evolving knowledge accumulation phase? 4.In Table 2, are OmniMed-Rad, OmniMed-Mic, and SLAKE separate adaptation settings or part of a continuous learning pipeline? 5.What is the exact task definition, and how are the reported metrics (formula) computed?

  • 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 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    My score is mainly driven by my main concerns listed above.

  • Reviewer confidence

    Very confident (4)

  • [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 regarding the experimental setting are addressed.



Author Feedback

We thank all reviewers. We first apologize for an unintentional model-name error: due to an internal API-provider mislabeling, the model we reported as Qwen2.5-VL-72B is in fact Qwen3-VL-235B-A22B (MoE). Since its activated parameters are smaller than a dense 72B, our actual API cost is even lower than the claim. All main conclusions are unaffected, and we will correct the related typos in the camera-ready. 1.Comparison with Standard RAG (R2, R3) We implemented a textbook-RAG baseline on OmniMedVQA-Rad: four open-access NCBI radiology textbooks (Chest/Heart/Vessels, Abdomen/Pelvis, Brain/Head-Neck/Spine, Medical Imaging Systems), chunked and BM25-retrieved into the same frozen backbone. Textbook-RAG drops Qwen2.5VL 67.21%->60.36% (-6.85) and Qwen3-VL 80.18%->77.09% (-3.09), while EvoMed gains +6.61 and +5.46 on identical backbones. Since the task needs precise visual judgement, imprecise textbook context distracts the model’s reasoning. EvoMed’s gain instead comes from compact, error-driven, task-adaptive clinical priors, not retrieval per se. 2.Fairness vs. SFT Medical VLMs (R3) We agree Table 2 is not same-backbone: LLaVA-Med/Med-Flamingo use 2023-era backbones with different training distributions, so part of the gap is backbone strength. Our controlled evidence is the same-backbone result: with only 500 evolution samples, EvoMed lifts the frozen 7B 67.21%->73.82% and the larger backbone 80.18%->85.64% on OmniMed-Rad. SFT medical VLMs are included to show distribution overfitting: LLaVA-Med reaches 85.34% on SLAKE (in-distribution) but collapses to 52.91% on OmniMed-Rad, below even the zero-shot generalist. Because EvoMed never updates backbone params, it avoids such over-specialization while still improving on the strong Qwen-VL backbone. We will moderate the claim and note backbone/training-data differences in Table 2.3.Evaluation Protocol and Data Splits (R2, R4) OmniMed-Rad is the radiology subset we filtered from OmniMedVQA; from it we randomly sampled a 500-example evolution set (Rad-500, MRI 306/CT 119/X-ray 75) and a disjoint 550-example test set (MRI 318/CT 153/X-ray 79), the two have no overlap. CT/X-ray/MRI in Table 1 are modality-stratified subsets of this test set, not sequential tasks; reported numbers are final accuracies after one KB built from Rad-500.In Table 2, OmniMed-Rad/Mic/SLAKE are independent adaptations, each with its own 500-sample evolution set and KB—not a continual-learning pipeline. Task is closed-set Med-VQA; acc=#correct/#test. 4.Semantic Triplet (R2) The triplet (Query Content, Modality, Organ/Tissue) is a small, well-defined schema. “Query Content” denotes the clinical intent of the question (e.g., presence, position, or diagnosis), not the answer itself. Because the schema is small and discrete, the LLM produces these triplets reliably across all our runs, so triplet noise is not a practical bottleneck. 5.Reliability of V_desc and Triplets (R2) Triplet generation is highly accurate due to its discrete schema. For open-ended V_desc, quality scales with backbone capacity: 235B-MoE yields noticeably more faithful descriptions than the 7B, consistent with its larger same-backbone gain and Table 3’s asymmetric transfer (the larger-model KB transfers better to the 7B than vice versa). The cross-encoder reranker further mitigates residual V_desc errors before generation. 6.LLM-as-Judge (R2) The LLM judge only decides whether the generated answer matches the closed-set ground-truth option—a discrete-matching task, not open-ended grading. We already manually verified all evaluation outputs align with human assessment. 7.Catastrophic Forgetting (R2) Since EvoMed keeps all backbone parameters frozen, it avoids parameter-update-induced forgetting by construction; we did not run a full continual-learning forgetting protocol. We will revise “completely circumvents catastrophic forgetting” to “avoids parameter-update-induced forgetting risk” and leave a formal forgetting-rate study to future work.



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.

    The paper proposes EvoMed, a training-free framework for adapting generalist VLMs to medical VQA by accumulating an external knowledge base through self-evaluation and retrieving relevant clinical priors during inference. Reviewers found the motivation timely and practically appealing, especially the avoidance of parameter updates, the reported data efficiency using a small evolution set, and the cross-model knowledge transfer analysis.

    However, the reviews raise several important concerns that need to be addressed in the rebuttal. First, reviewers noted that EvoMed is essentially a retrieval-augmented framework, but the paper does not compare or clearly distinguish it from simpler RAG alternatives. Please clarify, based on the current evidence, why the proposed self-evolving Actor/Critic/Summarizer/Updater loop is necessary and what supports the claim that it provides benefits beyond simply retrieving relevant medical context.

    Second, reviewers questioned the fairness of the comparison with fine-tuned medical VLMs, since models such as LLaVA-Med and Med-Flamingo use older/weaker backbones, while EvoMed builds on Qwen2.5-VL, whose zero-shot performance is already substantially stronger. Please clarify how these comparisons should be interpreted and whether the claims about outperforming fine-tuned medical VLMs should be moderated.

    Third, reviewers raised concerns about the evaluation protocol and task setup. Please clarify how the 500 evolution samples and the evaluated test subsets were selected, whether there is any overlap or leakage between evolution and test data, and whether OmniMed-Rad, OmniMed-Mic, and SLAKE are separate adaptation settings or part of a continual-learning sequence. Also, clarify whether CT, X-ray, and MRI in Table 1 are separate tasks or modality subsets.

    Finally, reviewers asked for clarification of key methodological details, including the definition of the “Query Content” field in the semantic triplet, the reliability of the generated visual descriptions and the retrieval triplets, the role of the LLM-as-judge in evaluation, and the basis for the claim that EvoMed circumvents catastrophic forgetting.

  • 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.

    I recommend acceptance. The rebuttal addressed the Meta-AC’s main concerns by adding a textbook-RAG baseline that hurts accuracy while EvoMed improves it on the same backbones, clarifying the evaluation protocol (disjoint 500 evolution and 550 test samples, modality subsets not sequential tasks, and independent adaptations across OmniMed-Rad, OmniMed-Mic, and SLAKE), defining the semantic triplet fields, and softening the catastrophic-forgetting claim. The authors also acknowledged that the comparison with older medical VLMs is not same-backbone and committed to moderating the claim, while pointing to same-backbone gains as the controlled evidence. R3 and R4 were convinced and either maintained or raised their scores to Accept, while R2 stayed at Reject mainly on positioning against other frozen-backbone RAG methods and V_desc faithfulness, which are not invalidations of the core results. The disclosed model-name error (the reported 72B model is actually a 235B MoE) is concerning but does not change the same-backbone improvements and can be corrected in the camera-ready.



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.

    Overall this is timely paper to adapt general VLM for medical VQAs. The authors were able to address the raised major concerns. This paper might raise good discussion at MICCAI. Hence, I vote for acceptance.



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 majority of the reviewers recommend acceptance following a solid rebuttal, which successfully demonstrated the distinct merits of the self-evolving mechanism through a rigorous textbook-RAG baseline ablation. While the authors disclosed a notable model mislabeling error (confounding a 235B-MoE with a 72B dense backbone), this does not invalidate the underlying framework or the controlled same-backbone performance gains. Consequently, the paper is recommended for acceptance, conditional on a thorough and transparent correction of all model specifications in the final camera-ready version.



back to top