acceptodds
Withdrawn from ICLR 2027

MedEvo: Medical Reasoning Evolution through Tag- Guided Navigation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Medical multimodal reasoning has advanced through teacher-generated rationale distillation, verifier- or reward-guided learning, and on-policy supervision over student-generated rollouts. Although these paradigms can exploit mistakes through rejection, reward, preference, process, or token-level signals, explicitly recovering valid reasoning states from failed medical trajectories and converting localized corrections into reusable, student-relative supervision remains underexplored. Moreover, general-domain correction mechanisms do not directly address medical failures grounded in modality-specific visual evidence, anatomical context, evidence integration, and differential diagnosis. We propose MedEvo, a failure-conditioned framework for evolving medical multimodal reasoning through localized trajectory repair and feedback-driven data evolution. At the trajectory level, MedEvo uses step-level uncertainty to propose a candidate failure region, employs a gold-blind auditor to identify a safe rollback boundary and generate a taxonomy-guided minimal recovery hint, and lets the original student preserve the valid prefix while regenerating only the erroneous suffix. Repair-aware learnability filtering then retains answer-correct, locally learnable trajectories that remain compatible with the student policy. At the data level, a fine-grained medical taxonomy characterizes checkpoint-specific capability gaps and guides the selection of targeted supervision for the next evolution round. Using supervised fine-tuning alone, MedEvo improves Qwen3.5-9B on MedXpertQA-MM from 47.40% to 54.50% and raises its seven-benchmark mean from 65.57% to 69.12%, a gain of 3.55 percentage points. Experiments on 9B, 35B, and 397B models show gains in mean benchmark accuracy across scales, while round-wise comparisons examine feedback-driven data evolution. These results establish localized failure repair and structured feedback as a practical step toward recursive self-improvement in medical reasoning.

Primary area:
generative models
Keywords:
Medical Multimodal Reasoning, Failure-Conditioned Data Synthesis, Learnability Filtering

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QrC5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a new iterative training method to optimize multimodal LLMs for medical reasoning. The core loop involves a student generating traces to medical problems, after which the traces are reviewed and tagged by the teacher to identify specific points in the traces which diverged from correct reasoning. The student then regenerates these tagged traces, keeping the correct prefix of the trace, and only regenerating the incorrect suffix. This results in higher quality student-generated traces for future supervision, which is combined with further evaluation on a feedback set to direct the model to identify the next best questions for future iterations of the loop. When applying these steps to finetune Qwen3.5-9B and other larger variants, the model significantly surpasses the base model's performance and outperforms other similarly sized competitor models.

Critical strengths

The method makes intuitive sense, and the authors present strong results to back it up. The recipe also generalizes across model sizes, which encourages follow ups with larger model size or increased rounds of iterations. I also found the paper well written and easy to follow the methodology.

Critical weaknesses

The paper mentions several baselines in the abstract, such as supervised finetuning, reinforcement learning, and on-policy distillation, but does not compare the performance of training with these methods to the results of the new proposed method. A comparison to these approaches would solidfy the contribution and show why the proposed method is better than using an existing approach. More concretely, these are some alternative approaches which I believe would be strong baselines to compare with:

  • Supervised finetuning on teacher-generated traces (obtained with rejection sampling using the final target answer as the correctness signal)
  • Reinforcement learning, with format reward and final answer correctness reward
  • On-policy distillation using a strong open-source teacher model

Another aspect of the method which raises some concern with me is the assumption that the teacher can identify incorrect reasoning in the student trace. This is a strong assumption, and even if it is true, suggests that using the teacher alone for direct distillation or creating teacher-generated traces would be a potentially stronger baseline.

Table 1 is a bit misleading. For example, in the <10B open source model section of the table, the entire MedEvo row is bolded, when LingShu actually achieves higher performance on SLAKE and PathVQA.

Questions for authors

Have you tried any ablations where you compare different existing approaches such as SFT, RL, OPD, on the same dataset you use in your new approach?

Do you assume the teacher is strong enough to identify incorrect medical reasoning in the student trace? Are there any instances where you observe inaccuracies in the correct reasoning cutoff point identified by the teacher model?

Have you tried your approach using the same base model used to train the other models you compared with (Qwen2.5-VL-Instruct for example)

Reviewer_8Bd6OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

MedEvo is a methods that ask LLM to learn from its own mistakes. When the model answers incorrectly, the system will find out the earliest mistakes in reasoning, and let the model keep the correct part in front and only redo the wrong part in the back. After the revised reasoning is filtered, training data will be added. After each round of training, the author selects the next round of exercises according to the weak items of the model. Experiments show that this methods can improve the score.

Critical strengths
  • The proposed method yields a considerable improvement to the base model.
  • Ablation studies showed that the different settings of the method are beneficial to the model's gains.
Critical weaknesses
  • The data construction pipeline is a conventional scheme. In medical LLM, HuatuoGPT-o1 (Chen, 2024) had proposed a relatively similar scheme. The construction of CoT data is achieved by observing the wrong conclusion judge model and constantly prompting the model to be modified. The main difference maybe the authors used the model's trajectory to refine it by itself, but not using an external model to generate the trajectory, but I can't find the ablation study related to this to support it is necessary.
  • Addressing the model's incorrect answers and repairing the corresponding incorrect answers to improve model performance is itself a widely used trick in LLM trianing.
  • Training data are not clearly indicated at all (I can't find a detailed info about the sources and filtering strategy of the 81,293 training samples). It's very important to show these details, cuz this determinees the reliability of GPT-based process supervision. As far as I know, the content of many current open-source datasets is not sufficient for GPT to judge all the generated content of the model (R0, R1...). For example, datasets such as PathVQA may only provide some simple Q&A. In this case, GPT may have a considerable bias when detecting mistakes generated by the model, as it relies entirely on GPT's own discriminative ability rather than strong evidence. I hope the authors can provide a relatively clear analysis and explanation on this point.
  • This paper is written very illogically. For example:
    • Regarding the definitions of R0, R1, R2, and R3, the authors completely fail to expain what these refer to in the experiments part. I can only find the definition after I carefully read all the formulas in Section 3. Such reading logic is inconvenient for readers.
    • I fail to understand the purpose of the later model scale-up experiments. The authors merely scaled up their own approach to 35B and 397B models, and they did not even bother to write the model names clearly.
Questions for authors

Please refer to the weaknesses.

I think the paper should be carefully designed and rewritten to fit the acceptance bar of ICLR.

If I have missed any method innovation in the details, please point it out to me, and I will apologize for that. But forgive me, I am really unwilling to give a high score to a paper with such a writing attitude.

Reviewer_R1zROpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies how to learn from failed medical reasoning trajectories while preserving the useful parts of the reasoning. Both 9B and 35B achieved the best average performance across several datasets compared to other baselines, but I have some concerns about the novelty and whether the experiments support the main claims.

Critical strengths

The framework uses model failures both to construct corrected training examples and to identify what the model should learn next. This provides a coherent approach to improving medical reasoning and is a potentially valuable contribution, even though several individual components build on existing methods.

Critical weaknesses
  • I am not sure the paper fully supports the claim "Moreover, general-domain correction mechanisms do not directly address medical failures grounded in modality-specific visual evidence, anatomical context, evidence integration, and differential diagnosis." in the abstract. The medical taxonomy and targeted hints address these issues in the method design, and the ablations provide some supporting evidence. However, it remains unclear whether they improve visual grounding, anatomical interpretation, evidence integration, and differential diagnosis beyond a general correction method given the same medical inputs.

  • The bolded values in Table 1 are not always the highest. e.g, Lingshu performs better than MedEvo on SLAKE and PathVQA. Please clarify what the bold indicates or reserve it for the best results within each comparison group. The discussion should also acknowledge where other models perform better.

  • Citations are needed in the introduction. The authors cite a lot of paper in the Related Work, but to support the claims in the Introduction, citations are also needed. The statements about most reasoning supervision learns from successful trajectories and discarding failed trajectories need references. The latter also seems too broad, since several existing methods explicitly learn from failures.

  • Methodological novelty. There is already substantial work on self-distillation, corrective hints, local repair, and learning from failed trajectories. HHD already uses hints to guide student-generated supervision, MOTAB reuses safe prefixes, and ROSD localizes errors and restricts distillation to the affected suffix. TAPO also constructs corrective trajectories with capability-dependent selection. I am not yet convinced that the paper clearly establishes what is new beyond combining these ideas in a medical setting. The medical taxonomy and feedback-guided data selection may be the main contribution, but this needs a more direct comparison and stronger experimental support. if those medical taxonomy and feedback-guided data selection are the main contribution, then results of how much those contributes to the performance improvement is needed.

  • The baseline comparisons need clarification.It is unclear whether the methods use comparable training rounds, generation budgets, and amounts of supervision.

  • MedEvo uses GPT-5.6-sol to identify errors and provide hints. The direct-distillation baseline uses Qwen-3.8-Max to generate training solutions. If MedEvo performs better, it cannot cleanly determine how much comes from its repair procedure and how much comes from differences in the teachers’ guidance.

  • A comparison with training only on correct student responses would help establish the value of repairing failures. Removing the learnability filter and comparing suffix-only with full-response training would further clarify which components contribute to the gains.

  • The case study is inconsistent with the method description and Figure 1. In the chest radiograph case (case 08), the student first states that the costophrenic angles are not blunted and that pleural effusion is unsupported. The hint appears after these statements, although the method says to truncate before the first consequential error. Were these statements retained in the continuation input or shown only for illustration? Please mark the rollback boundary and distinguish the discarded reasoning from the retained prefix.

  • The human audit raises questions about the quality of the supervision. Some boundaries are judged invalid, some repairs are not visually grounded, and some hints fail the leakage check. These findings should qualify the claims about preserving valid reasoning and providing answer-blind guidance. Please also clarify whether the audit examines candidates before filtering or the examples actually used for training, and provide the sample size and annotation procedure.

Questions for authors

I tried my best to read the appendix and found the appendix too long and difficult to follow. Please keep only the essential methodological details and supporting analyses, and clearly cross-reference them from the main text. Remove repetitive material and content that is not referenced in the main text.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.