Abstract

Reinforcement Fine-Tuning (RFT) has enabled medical Multimodal Large Language Models (MLLMs) to produce Chain-of-Thought (CoT) reasoning for visual question answering, yet these models suffer from confidence miscalibration—a systematic gap between expressed certainty and actual diagnostic accuracy that undermines clinical trust. We propose CARE, a Confidence-Aware medical REasoning framework that jointly optimizes accuracy and calibration through a dual-stage pipeline. First, a scalable Medical-CoT synthesis provides structured cold-start data for Supervised Fine-Tuning. Second, Group Relative Policy Optimization (GRPO) with a novel Confidence-Aware Reward (CAR) mechanism ties the model’s confidence to diagnostic correctness within the reward signal. Across three Medical VQA benchmarks, CARE achieves the highest diagnostic accuracy while obtaining the lowest Expected Calibration Error and Hallucination Rate, establishing a foundation for trustworthy clinical decision support.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6558_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/anotherbricki/CARE

Link to the Dataset(s)

N/A

BibTex

@InProceedings{DuYue_CARE_MICCAI2026,
        author = { Du, Yuetian AND Wang, Yucheng AND Chen, Zhenyuan AND Chen, Luyuan AND Zhang, Rongyu AND Zhang, Jinjian AND Zhou, Wei AND Xu, Zhijie AND Kong, Ming AND Zhou, Zhan AND Liu, Jie AND Zhu, Qiang},
        title = { { CARE: Confidence-Aware Reasoning for Reliable Medical VQA } },
        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 paper presents CARE, a two-stage post-training framework for medical VQA that targets both diagnostic performance and confidence calibration. The method first constructs label-consistent medical chain-of-thought trajectories and then trains a Qwen2.5-VL-7B backbone via SFT followed by GRPO using a Confidence-Aware Reward (CAR). CAR defines confidence from the model’s own token probabilities and rewards high confidence on correct outputs while penalizing high confidence on incorrect ones. Experiments on VQA-RAD, SLAKE, and PathVQA evaluate accuracy/recall, ECE, and hallucination rate against several recent medical reasoning MLLM baselines.

  • 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 paper addresses a clinically meaningful problem that is more specific than simply improving Med-VQA accuracy. Focusing on whether model confidence tracks diagnostic correctness is important for triage, escalation, and safe downstream use, and this is a well-motivated target for medical reasoning models.

    The proposed training recipe is simple and practically implementable. CAR is integrated directly into GRPO without introducing a separate uncertainty head or a complicated auxiliary model, which makes the method easy to understand and potentially easy to adopt in related medical MLLM pipelines. The empirical results are strong and consistent on the reported benchmarks. In Table 1, CARE is the only method reported to rank first across accuracy, ECE, and hallucination rate on all three datasets, which suggests that the method is not merely trading calibration for accuracy or vice versa.

    The training-stage analysis is useful. The separation of Training-Free, SFT, RL, and SFT+RL provides a more informative picture than a single end-point comparison, and the open-ended versus closed-ended discussion offers practical insight into when structured CoT initialization appears most helpful.

  • 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 central contribution of CAR is not cleanly isolated by the current ablations.

    Table 2 varies training stages, but there is no same-backbone comparison between RL with a standard accuracy/form reward and RL with CAR.

    The evaluation metrics could be removed to set aside more room for ablation studies.

  • 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?

    The paper is well organized, the practical motivation is clear, and the headline empirical numbers are strong across three benchmarks. It could be better if the ablation studies provides more evidence of the performance gain brought by CAR.

  • 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.

    The author response can address my concerns.



Review #2

  • Please describe the contribution of the paper

    The paper proposes CARE, a confidence-aware reinforcement learning framework for medical VQA. It introduces a confidence-based reward to align model confidence with diagnostic correctness. The method combines CoT data synthesis, SFT initialization, and RL optimization. Experiments show improvements in accuracy, calibration, and hallucination across multiple benchmarks.

  • 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.

    Strengths 1.The paper studies confidence miscalibration in medical VQA. This is an important and practical problem for clinical deployment. 2.The pipeline is clear and well structured. The combination of CoT data synthesis, SFT initialization, and RL optimization is easy to follow. 3.The experiments cover multiple datasets. The evaluation includes accuracy, calibration, and hallucination metrics. Results are consistently strong across benchmarks. 4.The overall presentation is good 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.

    Weaknesses

    Method 1.The paper defines confidence as the average token probability of the generated output. This definition is explicit, but is it a valid proxy for answer-level confidence? It seems closer to generation likelihood than epistemic uncertainty. How sensitive is this measure to output length or decoding settings such as temperature? 2.The method may lack sufficient insight. Several recent works [1,2,3] have explored confidence-aware or confidence-driven optimization. What is the key methodological difference between CAR and these approaches? 3.The reward design may not guarantee true calibration. It encourages higher confidence scores derived from token-level probabilities on correct outputs and penalizes them on incorrect ones. But does this ensure alignment with empirical accuracy, or only reshape token-level likelihood?

    Experiments 1.The hallucination metric relies on a VLM judge (Lingshu-32B), but the inference setup is unclear. How are temperature, prompting, and evaluation consistency controlled? 2.The contribution of the proposed reward is not isolated. In Table 2, the ablation only compares training stages, without isolating CAR. It may be helpful to include comparisons with RL without CAR to better understand its impact. 3.It may also be important to report statistical significance. In Table 1, results on small datasets such as VQA-RAD are presented without variance. How stable are these improvements across different runs?

    [1] Gu, Bowen, et al. “Probabilistic medical predictions of large language models.” npj Digital Medicine 7.1 (2024): 367.[2] Li, Pengyi, et al. “Confidence is all you need: Few-shot rl fine-tuning of language models.” arXiv preprint arXiv:2506.06395 (2025). [3] Prabhudesai, Mihir, et al. “Maximizing confidence alone improves reasoning.” arXiv preprint arXiv:2505.22660 (2025).

  • 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 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?

    The paper studies an important problem and presents a clear pipeline with strong empirical results across multiple benchmarks. However, the method is not fully convincing. The confidence definition based on token-level probabilities is not well justified, and the contribution of the proposed reward is not clearly isolated.

  • 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.

    Thank you for the clarification on the confidence computation, the role of CAR, and the evaluation settings. These responses address most of my concerns. The hallucination evaluation setup is clearer, though the reliability of the judge could still be better validated in future work (e.g., consistency across multiple evaluation rounds). Overall, I find the rebuttal convincing and will support acceptance.



Review #3

  • Please describe the contribution of the paper

    The paper introduces an automated pipeline for constructing reasoning trajectories from visual content, questions, and answers, along with a novel reward mechanism designed to calibrate the output token confidence of the reasoning model, which encourages high confidence for reasoning paths that lead to correct answers and low confidence for those that do not.

  • 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 research problem addressed in this paper, calibrating the output confidence of MLLMs, is both practical and significant. Well-calibrated confidence scores can serve as a critical indicator for clinicians, enabling them to better assess the reliability of model outputs and make more informed decisions.

  • 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 connection between the two proposed contributions needs to be clarified. Specifically, it remains unclear how the constructed CoT data benefits the confidence-calibrative RL training, or conversely, how the proposed RL reward leverages the constructed CoT data to achieve better calibration. Further elaboration on the interplay between these two components would help readers understand how they jointly strengthen the overall framework.

    2.The novelty of the proposed CoT curation process warrants further clarification. Several existing methods already construct CoT data from visual content, questions, and final conclusions [1], making the distinction of the proposed approach less apparent. Including a discussion of related multi-modal medical CoT construction methods (e.g., [1][2]) would provide important context and help readers better appreciate the specific contributions of this work.

    3.The rationale behind the confidence calibration reward requires more justification. The reward is computed by averaging the confidence scores across all output tokens; however, not all tokens are equally informative with respect to the final conclusion — for instance, tokens such as or punctuation marks carry little semantic relevance. Whether the inclusion of such tokens distorts the reward signal is not discussed. Additionally, the claim that "a white-box signal derived from the model's own output distribution can be directly tied to diagnostic correctness within the reward function" lacks sufficient supporting evidence and should be substantiated more rigorously.

    4.Additional experiments are needed to more convincingly support the paper’s claims. The rationale for benchmarking compact-scale models (≤3B parameters) against the trained 7B model should be clarified, as this comparison may not be entirely fair. More importantly, ablative studies that individually quantify the contributions of the proposed CoT construction and the new reward to overall performance improvement are necessary. The current ablation focuses on different training schemes, which does not directly validate the two core contributions. Furthermore, the confidence shift shown in Figure 3 is somewhat expected given that the reward directly supervises this behavior. A more informative analysis would present the confidence shift conditioned separately on correct and incorrect conclusions, which would provide stronger evidence that the model genuinely learns to re-calibrate its confidence during the reasoning process rather than simply responding to the reward signal.

    [1] Niu, Shuai, et al. “Multimodal clinical reasoning through knowledge-augmented rationale generation.” arXiv preprint arXiv:2411.07611 (2024). [2] Sun, Haoran, et al. “Chiron-o1: Igniting multimodal large language models towards generalizable medical reasoning via mentor-intern collaborative search.” arXiv preprint arXiv:2506.16962 (2025).

  • 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.Figure 2 contains display errors in the Confidence-Aware Reward section that should be corrected before publication.

    2.The confidence allocation depicted for the “Wrong Diagnosis” case in Figure 1 may be misleading to readers. The model appears to exhibit a low confidence distribution when producing an incorrect answer, which could actually be interpreted as a desirable behavior, since outputting a wrong answer with low confidence is generally preferable to doing so with high confidence. Clarifying the intended message of this illustration, and distinguishing between what the current model does versus what the proposed method aims to achieve, would help avoid confusion.

  • 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.

    (2) Reject — should be rejected, independent of rebuttal

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

    The unclear connection between the two proposed contributions, the insufficient justification for the confidence calibration reward, the limited novelty of the CoT data curation pipeline, and the lack of ablative experiments that directly validate the core claims, collectively prevent me from recommending acceptance at this stage. Substantial revisions addressing these issues would be necessary before the paper is ready for submission to another venue.

  • 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.

    I thank the authors for their efforts. The rebuttal adequately addressed my concerns regarding the confidence calibration reward design, which should be reflected in the revision. However, my concerns about the CoT construction method remain unresolved. In the rebuttal, the authors stated that “the Medical-CoT synthesis is not claimed as a standalone novelty over prior medical rationale construction.” Yet in the original manuscript, the CoT construction method is listed as a main contribution, occupies roughly half of the main figure, and is described in a dedicated subsection. This inconsistency leaves the concerns raised in Points 1 and 2 of my original review unaddressed. Additionally, the authors introduced new experimental results directly in the rebuttal, and it is unclear whether these can be fully incorporated and discussed in the revision within the page limit. For these reasons, I am moving my rating to Weak Reject, but still leaning towards rejection.



Author Feedback

We thank the AC and reviewers for recognizing the clinical relevance, clear story, and comprehensive evaluation of CARE. The major concerns are whether CoT/CAR contributions are isolated, confidence validity, and how CARE differs from related work. We address them below.

Ablation for CoT and CAR (R1Q1/R1Q2/R2Q5/R4Q4). We conducted an ablation study under different GRPO settings on SLAKE: Setting / ACC / ECE / HR Base GRPO: 71.25 / 25.92 / 16.37 with CoT: 77.94 / 18.89 / 14.88 with CAR: 82.47 / 14.83 / 12.38 CARE: 87.30 / 11.50 / 7.01 This confirms their complementarity: CoT strengthens reasoning initialization, while CAR improves calibration beyond standard GRPO.

Validity of confidence proxy (R2Q1/R4Q3/R2Q3). Prior work defines calibration as confidence-performance alignment [1]. Accordingly, we use ECE to directly evaluate whether our answer-level confidence proxy aligns with actual accuracy. The consistent ECE reductions in Table 1 and our new ablation support the validity of this proxy, showing that CAR improves confidence-accuracy alignment rather than merely increasing likelihood.

Differentiation from confidence-driven works (R2Q2). Mentioned rencent work ask the model to report confidence, or reward high confidence alone. Instead, CAR use confidence in GRPO as a correctness-conditioned calibration term.

Does CAR really recalibrate? (R2Q3/R4Q4). We further added conditioned confidence analysis on VQA-RAD. After CAR, confidence increases on correct predictions and decreases on incorrect predictions: OPEN Question: correct +0.84, incorrect -0.59 CLOSED Question: correct +0.07, incorrect -0.16 This shows CARE shifts confidence according to correctness.

Statistics (R2Q6/R4Q4). We added statistical analysis on VQA-RAD (4-times): Model / ACC / ECE MedVLThinker-7B: 65.44±0.68 / 45.95±0.53 Lingshu-7B: 65.99±2.21 / 40.51±1.33 CARE: 76.32±0.95 / 20.31±1.28 CARE remains clearly better than strong 7B baselines in both accuracy and calibration, supporting the stability of the improvement on this small dataset.

Differentiation from medical CoT works (R4Q1/R4Q2). The Medical-CoT synthesis is not claimed as a standalone novelty over prior medical rationale construction. Its role is to provide a cold start for confidence-aware RL. CoT stabilizes the reasoning format and answer span, reduces noisy exploration in open-ended medical reasoning, and supplies structured trajectories whose final answers can be verified. Thus, CoT improves reasoning structure, then CAR aligns answer confidence with correctness.

Confidence computation in CARE (R4Q3). We compute confidence only within the span, excluding reasoning, formatting, and special tokens (like ). The claim mentioned in your question can be support by ECE metrics in our experiment, detailed explanation please refer to the second response question. This is also consistent with recent confidence-based reasoning studies [2].

Evaluation details and fairness (R1Q3/R2Q4/R4Q4/R4Q5/R4Q6). We will add the full hallucination-judge protocol: Lingshu-32B uses deterministic decoding, a fixed prompt/rubric, identical inputs for all methods, so random seeds do not affect the judge outputs. Compact models are included only for completeness; the primary comparison is with 7B-8B models, where CARE still leads. We will also fix the Fig. 2 rendering issue, revise Fig. 1 to distinguish existing miscalibration from CARE’s target behavior, and remove metric details to save space for ablations.

We believe these new ablations, conditioned confidence analysis, and statistical results directly address the main concerns and substantially strengthen the evidence for CARE.

[1] Taming Overconfidence in LLMs: Reward Calibration in RLHF, ICLR 2025.[2] Confidence as a Reward: Transforming LLMs into Reward Models, ICLR 2025.




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.

    As the AC, personally I like this paper and I believe my judgement is also aligned with the reviewers. The strength of this work is clear, e.g. topic relevance, well-structured story and the comprehensive evaluation.

    I invite the authors to rebuttal and would like to see if the following potential weaknesses can be addressed: 1.Explanation about the validity of the answer-level confidence proxy 2.The methodology differentation to the similar works as mentioned by R2 and R4 3.The persuasiveness of the evaluation, e.g. statistics analysis, the contribution of the proposed CoT construction and reward design

    Good luck!

  • 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 authors offered a detailed rebuttal and several concerns are well addressed. I recommend the acceptance of the paper, but the authors shall not view this work as the final product of the project. If you want to publish at a higher platform, you may also want to address the weaknesses of using generated CoT supervision and judge-based hallucination scoring. Best luck and looking forward to seeing your future work.



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.

    As noted by the reviewers, many raised concerns have been properly addressed in the rebuttal. A remaining point of dispute concerns the curation of SFT cold-start data. The meta-reviewer partly agrees with R3’s argument that, as CoT synthesis for cold start is a common practice, the authors should not overclaim their contribution in this regard in the camera-ready version.



back to top