Pith. sign in

REVIEW 3 major objections 5 minor 22 references

CARE: Confidence-Aware Reasoning for Reliable Medical VQA

T0 review · 3 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read CARE claims that a correctness-conditioned confidence reward lets one 7B medical VQA model lead all benchmarks in accuracy, calibration, and hallucination rate.

desk verdict The confidence-aware reward is a real idea, but the paper's headline numbers may be memorization because the train/test split is never disclosed. read the letter →

arxiv 2608.10964 v1 pith:YJ3GTJ5Q submitted 2026-08-11 cs.CV cs.AI

classification cs.CVcs.AI
keywords medicalvisualquestionansweringconfidencecalibrationreinforcementfine-tuninggrouprelativepolicyoptimizationchain-of-thoughtexpectederrorhallucinationmultimodallargelanguagemodels
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Confidence-aware medical reasoning (CARE) is a training recipe for medical visual question answering that claims to remove the usual trade-off between accuracy and calibration. The paper argues that the reason current reinforcement-fine-tuned medical models are untrustworthy is not lack of reasoning but a gap between expressed confidence and actual correctness. It therefore builds a two-stage pipeline: a scalable synthesis of verified chain-of-thought traces for supervised fine-tuning, then GRPO with a reward that rewards confident correct answers and penalizes overconfident wrong ones. On VQA-RAD, SLAKE, and PathVQA, the trained 7B model reports the highest accuracy together with the lowest expected calibration error and hallucination rate among the compared models. If the claim holds, confidence can be optimized as a first-class training objective rather than measured after the fact.

What carries the argument

The Confidence-Aware Reward (CAR) is the object that carries the argument: it computes answer-level confidence $C(a_i) = \frac{1}{|a_i|} \sum_{j=1}^{|a_i|} \pi_\theta(t_j \mid V, Q, t_{<j})$ over only the tokens inside the `<answer>` span, then combines it with the correctness indicator through $R_{\mathrm{calib}} = R_{\mathrm{out}} C(a_i) - \lambda (1-R_{\mathrm{out}}) C(a_i)$. CAR plugs into the GRPO advantage estimator, replacing scalar accuracy rewards with a composite reward $R_{\mathrm{form}} + R_{\mathrm{out}} + R_{\mathrm{calib}}$. Its role is to make calibration a differentiable, optimizable target inside policy gradient, while the auxiliary Medical-CoT synthesis provides the structured cold start that makes answer tokens extractable and verifiable.

What would settle it

Resolve the train/eval overlap question: enumerate the image-question pairs used to build DCoT, intersect them with the evaluation sets, and rerun CARE on the disjoint portion. If any benchmark query was used for training, or if the reported gains shrink on a clean split, the central claim of simultaneous improvement is not established.

Watch

Extended reading notes

Core claim

The paper's central discovery is that tying the model's confidence to its correctness in the reward signal lets a single 7B model win on accuracy, calibration, and hallucination simultaneously, where every baseline trades one off against another. The key result is Eq. (6): for each candidate answer, confidence $C(a_i)$ is the mean probability of the tokens in the answer span, and the calibration reward is $R_{\mathrm{calib}} = R_{\mathrm{out}} C(a_i) - \lambda (1-R_{\mathrm{out}}) C(a_i)$ with $\lambda = 0.5$. A correct high-confidence answer is rewarded; an incorrect high-confidence answer is penalized. Inside GRPO, this composite reward is normalized into group-relative advantages, so the policy learns to say "I am confident" only when it is right. Reported numbers across VQA-RAD, SLAKE, and PathVQA show accuracy 0.767, 0.873, and 0.689, ECE 0.202, 0.115, and 0.290, and hallucination rates 0.048, 0.070, and 0.059, all best among the compared models. Ablations further show that closed-ended questions improve most from RL alone while open-ended questions need the SFT cold start.

Load-bearing premise

The load-bearing premise is that the accuracy and calibration numbers come from held-out test questions, not from the same VQA-RAD, SLAKE, and PathVQA instances that were used to synthesize the training reasoning trajectories.

Editorial extensions

If this is right

  • A single training objective can lift accuracy and calibration together, contradicting the usual observation that strong medical reasoning models sacrifice one for the other.
  • Open-ended diagnostic questions rely on the synthesized chain-of-thought cold start; without it, RL alone does not reach the same open-ended accuracy.
  • Closed-ended questions benefit most from the RL phase alone, where the constrained answer space lets confidence-aware exploration converge without SFT initialization.
  • Reasoning traces generated backward from the ground-truth answer and then filtered by a verifier reduce hallucination at the data-construction stage.
  • The framework's reward design needs only a verifiable correctness signal, so it is portable to any medical VQA benchmark with ground-truth answers.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same conditional calibration reward could transfer to any verifiable question-answering setting outside medicine, such as general science or legal QA, whenever a correctness check is available; the paper sketches this direction in its conclusion but does not test it.
  • The answer-span average token probability is one possible confidence estimator; replacing it with a calibrated variant, such as sampling-based agreement or verbalized confidence, would test whether CARE's gains are tied to this specific definition.
  • The hallucination metric relies on a judge model's normalized score, so a natural follow-up experiment is to have expert clinicians rate the same outputs and tell whether the lower hallucination rate reflects factual grounding rather than adherence to a preferred output format.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes CARE, a two-stage medical VQA training framework that combines synthetic Medical-CoT data for supervised fine-tuning with GRPO reinforcement learning using a Confidence-Aware Reward (CAR). The reward is intended to align answer-level confidence with diagnostic correctness: correct answers with high confidence are rewarded, while incorrect answers with high confidence are penalized. Experiments on VQA-RAD, SLAKE, and PathVQA report that CARE-7B achieves the best diagnostic accuracy, the lowest Expected Calibration Error, and the lowest Hallucination Rate among the compared medical reasoning MLLMs, including Med-R1, MedVLM-R1, MedMO, Lingshu-7B, MedVLThinker, and Fleming-VL. The central claim is that accuracy and calibration can be jointly improved without a trade-off.

Significance. If the reported results are valid, CARE would be a practically useful contribution: it demonstrates a concrete reward design for reducing confidence miscalibration in medical VQA, a safety-relevant issue that existing reasoning-focused medical MLLMs do not directly address. The paper includes a reasonable ablation of the two training stages, and the code is promised publicly. The claimed improvements are large and consistent across three benchmark datasets, which makes the result potentially important for the medical multimodal LLM community. However, the validity of the central claim rests on two load-bearing assumptions that are not currently established in the manuscript: that evaluation is performed on held-out test samples disjoint from the training data used to synthesize DCoT, and that the reported ECE reduction reflects a genuine improvement in calibration rather than an artifact of optimizing the same confidence measure used for evaluation. The method itself is simple and reproducible, and the weaknesses identified here appear addressable with additional experiments and protocol disclosure rather than being fatal to the approach.

major comments (3)
  1. [Section 2.2, Eq. (1) and Section 3.1, Table 1] The source dataset D_VQA is never concretely identified, and the paper never states whether DCoT was synthesized from training splits or from the full VQA-RAD, SLAKE, and PathVQA benchmarks. Since DCoT is used for both SFT and GRPO and the same benchmarks are used for evaluation, the accuracy, ECE, and HR gains in Table 1 could reflect memorization of ground-truth answers and answer-specific confidence patterns rather than held-out generalization. Please specify the exact datasets used for synthesis, disclose the train/test split for each benchmark, and confirm that no evaluation sample appears in either training phase.
  2. [Section 2.4, Eqs. (5)-(7)] The reported ECE improvement is substantially by construction. Eq. (5) defines answer-level confidence C(a_i), Eq. (6) uses C(a_i) to reward confident correct predictions and penalize confident incorrect predictions, and Eq. (7) computes ECE from the same confidence measure. Consequently, the training objective directly minimizes the quantity used for evaluation. To support the calibration claim, the paper should provide an independent test of calibration, for example by evaluating ECE with temperature scaling, a held-out confidence estimator, or token-level probabilities that are not part of the CAR objective.
  3. [Section 3.1, Eqs. (7)-(8), Table 1] The hallucination rate metric depends entirely on a single VLM judge, Lingshu-32B, with no validation of that judge against human annotation or against the ground-truth answers; the HR formula in Eq. (8) is also not fully specified because the normalization constant C and the role of the token-count weighting t_i are not defined precisely. Since Table 1 reports HR as a headline result, please either validate the judge's reliability on a sample of human-labeled trajectories or report simpler raw hallucination rates in addition to the weighted score.
minor comments (5)
  1. [Section 2.2] The notation D_VQA is introduced as "a standard Medical-VQA dataset," but the actual dataset name is never given; please state which dataset or datasets were used for Medical-CoT synthesis and clarify the relation to the three evaluation benchmarks.
  2. [Section 2.4] The format reward R_form is mentioned but never formally defined; please specify exactly how the <think> and <answer> delimiters are checked and how this reward is applied to open-ended answers with variable lengths.
  3. [Section 3.1] The open-ended Recall metric is not defined; please describe how coverage of the ground-truth answer Y within the predicted answer a_i is computed and how this interacts with the exact-match indicator for closed-ended questions.
  4. [Section 3.3, Figure 3] The confidence histograms in Figure 3 lack axis labels and units, and it is not clear whether the "pre" and "post" distributions come from the exact checkpoints used in Table 1; please clarify the checkpoint selection and make the figure self-contained.
  5. [Eq. (8)] The formula for HR contains an apparent formatting artifact ("C E h") and the hat notation is unexplained; please restate the equation cleanly and define all symbols.

Circularity Check

1 steps flagged · score 6.0 of 10

The headline ECE gain is the training objective by construction: the confidence C(a_i) optimized in Eq. (6) is the same quantity binned by Eq. (7), so the reported calibration improvement is not independent evidence; no train/test split is disclosed.

  1. fitted input called prediction [Section 2.4 (Eqs. 5-6) and Section 3.1 (Eq. 7)]
    "The calibration reward ties this answer-level confidence to diagnostic correctness: Rcalib(oi, Y) = Rout · C(ai) − λ(1 − Rout) · C(ai) ... This correctness-conditioned design differs from simply encouraging high confidence, and directly optimizes the confidence-accuracy alignment measured by ECE."

    Eq. (5) defines C(ai) as the mean token probability of the answer span. Eq. (6) rewards high C(ai) when the answer is correct and penalizes high C(ai) when it is wrong. Eq. (7) defines ECE by binning 'confidence scores' — the same C(ai) measure — against accuracy. Thus the reported ECE reduction is a direct aggregate of the objective being optimized during GRPO, not an emergent or independent property. The paper itself says CAR 'directly optimizes the confidence-accuracy alignment measured by ECE,' yet Section 3.3 presents the ECE reduction as confirmation of CAR's effectiveness.

full rationale

The only substantial circularity is in the calibration claim. The confidence estimate C(ai) defined in Eq. (5) is used both as the training signal in the CAR reward (Eq. 6) and as the 'confidence' that ECE bins in Eq. (7). Therefore, reporting lower ECE is essentially reporting success on the training objective, especially because no train/test split is disclosed for the DCoT corpus synthesized in Section 2.2 relative to the VQA-RAD, SLAKE, and PathVQA evaluation sets. This makes the central calibration part of the claim partially circular. The accuracy and hallucination-rate results are grounded in external metrics (answer matching and a VLM judge), so they are not circular by construction, although their validity depends on the undisclosed split. The self-citation [5] in the introduction is used only to motivate the well-known miscalibration problem and is corroborated by Figure 1 and external references, so it is not load-bearing. Overall, one of the three headline metrics reduces to the training objective, warranting a partial-circularity score of 6.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claims rest on several modeling and protocol assumptions. Lambda is the only explicitly tuned free parameter that directly enters the core reward equation. The confidence proxy, the unstated evaluation split, and the LLM-based verification and judging are domain assumptions whose failure would change the conclusions. No new physical or ontological entities are introduced; CAR is a reward function and Medical-CoT is a data pipeline, both internal to the training procedure.

free parameters (1)
  • lambda (calibration penalty) = 0.5
    Chosen by hand in Sections 2.4 and 3.2; controls how strongly overconfident incorrect answers are penalized. No sensitivity analysis is reported, so the ECE gains in Table 1 are not shown to be robust to this choice.
assumptions (4)
  • domain assumption Average token probability over the answer span is a valid measure of the model's confidence for calibration purposes.
    Defined in Eq. 5 and used in both Rcalib and the ECE metric. If this quantity is not the clinically meaningful confidence, the calibration claim is weakened.
  • domain assumption VQA-RAD, SLAKE, and PathVQA are evaluated on held-out splits, not on data used to synthesize DCoT or train the model.
    Section 3.1 names the datasets and Section 2.2 builds DCoT from them, but no split is described. The core table numbers depend on this premise.
  • domain assumption GPT-4o trajectory verification and the Lingshu-32B hallucination judge are sufficiently reliable and neutral to support filtering and cross-model HR comparison.
    Sections 2.2 and 3.1 introduce these LLM judges without validation against human raters or inter-judge agreement.
  • standard math GRPO with K=4 rollouts and the stated hyperparameters is a stable optimization setup for this task.
    The GRPO objective (Eq. 3) is standard; K=4, batch size 2, and bf16 are given in Section 3.2, but no reward scaling or seed details are provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of CARE: Confidence-Aware Reasoning for Reliable Medical VQA." pith.science (2026). https://pith.science/paper/YJ3GTJ5Q

@misc{pith2026260810964,
  author       = {Pith},
  title        = {Pith review of: CARE: Confidence-Aware Reasoning for Reliable Medical VQA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YJ3GTJ5Q}},
  note         = {Machine review of arXiv:2608.10964}
}
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 $\textit{confidence miscalibration}$---a systematic gap between expressed certainty and actual diagnostic accuracy that undermines clinical trust. We propose $\textbf{CARE}$, a $\textbf{C}$onfidence-$\textbf{A}$ware medical $\textbf{RE}$asoning 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 $\textbf{Confidence-Aware Reward (CAR)}$ mechanism ties the model's confidence to diagnostic correctness within the reward signal. Across three Medical VQA benchmarks, $\textbf{CARE}$ achieves the highest diagnostic accuracy while obtaining the lowest Expected Calibration Error and Hallucination Rate, establishing a foundation for trustworthy clinical decision support. Our code is available at https://github.com/anotherbricki/CARE.

Figures

Figures reproduced from arXiv: 2608.10964 by the authors.

Figure 1
Figure 1. Comparative Analysis of Diagnostic Reasoning Paths. Existing medical reason￾ing MLLMs such as MedVLM-R1 exhibit confidence miscalibration, where expressed certainty fails to reflect actual diagnostic accuracy. CARE explicitly aligns subjective confidence with diagnostic correctness, producing verifiable and interpretable reasoning trajectories for clinical decision-making. Reinforcement Fine-Tuning (RFT), where RL a… view at source ↗
Figure 2
Figure 2. Overview of the CARE Framework. The pipeline consists of two phases: (1) Medical-CoT Data Construction for synthesizing structured diagnostic reasoning paths, and (2) Two-Stage Optimization, comprising SFT cold-start and GRPO with the proposed Confidence-Aware Reward (CAR) mechanism to align confidence with diagnostic accuracy. 2.3 Two-Stage Optimization of CARE Leveraging the curated DCoT, CARE is trained in two st… view at source ↗
Figure 3
Figure 3. Confidence distributional shift. Red and green bars represent the normalized confidence scores pre- and post-training, respectively. After training, the mean shifts rightward with increased magnitude, reflecting improved confidence-accuracy alignment. to converge efficiently without SFT initialization. For open-ended questions, however, SFT+RL consistently dominates (e.g., 0.881 on SLAKE Open, 0.421 on PathVQA Open)… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 5 canonical work pages

  1. [1]

    Bai, S., et al.: Qwen2.5-vl technical report (2025), https://arxiv.org/abs/2502.13923

  2. [2]

    Chu, T., Zhai, Y., Yang, J., Tong, S., Xie, S., Schuurmans, D., Le, Q.V., Levine, S., Ma, Y.: Sft memorizes, rl generalizes: A comparative study of foundation model post-training (2025), https://arxiv.org/abs/2501.17161

  3. [3]

    DeepSeek-AI, et al.: Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning (2025), https://arxiv.org/abs/2501.12948

  4. [4]

    Deria, A., Kumar, K., Dukre, A.M., Segal, E., Khan, S., Razzak, I.: Medmo: Grounding and understanding multimodal large language model for medical images (2026), https://arxiv.org/abs/2602.06965

  5. [5]

    In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025

    Du, Y., Wang, Y., Kong, M., Liang, T., Long, Q., Chen, B., Zhu, Q.: Confidence calibration for multimodal llms: An empirical study through medical vqa. In: proceedings of Medical Image Computing and Computer Assisted Intervention – MICCAI 2025. vol. LNCS 15965. Springer Nature Switzerland (September 2025)

  6. [6]

    In: International Conference on Machine Learning

    Guo, C., Pleiss, G., Sun, Y., et al.: On calibration of modern neural networks. In: International Conference on Machine Learning. pp. 1321–1330. PMLR (2017)

  7. [7]

    He, X., Zhang, Y., Mou, L., Xing, E., Xie, P.: Pathvqa: 30000+ questions for medical visual question answering (2020), https://arxiv.org/abs/2003.10286

  8. [8]

    Huang, X., Wu, J., Liu, H., Tang, X., Zhou, Y.: Medvlthinker: Simple baselines for multimodal medical reasoning (2025), https://arxiv.org/abs/2508.02669

Show all 22 references
  1. [9]

    org/abs/2503.13939

    Lai, Y., Zhong, J., Li, M., Zhao, S., Yang, X.: Med-r1: Reinforcement learning for generalizable medical reasoning in vision-language models (2025), https://arxiv. org/abs/2503.13939

  2. [10]

    Lambert, N., et al.: Tulu 3: Pushing frontiers in open language model post-training (2025), https://arxiv.org/abs/2411.15124

  3. [11]

    Scientific Data5(180251) (2018)

    Lau, J., Gayen, S., Ben Abacha, A., et al.: A dataset of clinically generated visual questions and answers about radiology images. Scientific Data5(180251) (2018)

  4. [12]

    In: Advances in Neural Information Processing Systems

    Li, C., Wong, C., Zhang, S., et al.: LLaVA-Med: Training a large language-and- vision assistant for biomedicine in one day. In: Advances in Neural Information Processing Systems. vol. 36 (2024) 10 Y. Du et al

  5. [13]

    Li, Z.Z., et al.: From system 1 to system 2: A survey of reasoning large language models (2025), https://arxiv.org/abs/2502.17419

  6. [14]

    In: IEEE 18th International Symposium on Biomedical Imaging

    Liu, B., Zhan, L.M., Xu, L., et al.: SLAKE: A semantically-labeled knowledge- enhanced dataset for medical visual question answering. In: IEEE 18th International Symposium on Biomedical Imaging. pp. 1650–1654 (2021)

  7. [15]

    arXiv preprint (2025), https://arxiv

    Pan, J., Liu, C., Wu, J., Liu, F., Zhu, J., Li, H.B., Chen, C., Cheng, O., Rueckert, D.: MedVLM-R1: Incentivizing medical reasoning capability of vision-language models (VLMs) via reinforcement learning. arXiv preprint (2025), https://arxiv. org/abs/2502.19634, arXiv:2502.19634

  8. [16]

    Shao, Z., et al.: Deepseekmath: Pushing the limits of mathematical reasoning in open language models (2024), https://arxiv.org/abs/2402.03300

  9. [17]

    Shu, Y., Liu, C., Chen, R., Li, D., Dai, B.: Fleming-vl: Towards universal medical visual reasoning with multimodal llms (2025), https://arxiv.org/abs/2511.00916

  10. [18]

    Sun, H., Jiang, Y., Lou, W., Zhang, Y., Li, W., Wang, L., Liu, M., Liu, L., Wang, X.: Chiron-o1: Igniting multimodal large language models towards generalizable medical reasoning via mentor-intern collaborative search (2025), https://arxiv.org/ abs/2506.16962

  11. [19]

    Team, L., Xu, W., Chan, H.P., Li, L., Aljunied, M., Yuan, R., Wang, J., Xiao, C., Chen, G., Liu, C., Li, Z., Sun, Y., Shen, J., Wang, C., Tan, J., Zhao, D., Xu, T., Zhang, H., Rong, Y.: Lingshu: A generalist foundation model for unified multimodal medical understanding and rea...

  12. [20]

    arXiv preprint (2023), https://arxiv.org/abs/2305.14975, arXiv:2305.14975

    Tian, K., Mitchell, E., Zhou, A., et al.: Just ask for calibration: Strategies for eliciting calibrated confidence scores from language models fine-tuned with human feedback. arXiv preprint (2023), https://arxiv.org/abs/2305.14975, arXiv:2305.14975

  13. [21]

    arXiv preprint (2023), https://arxiv

    Xiong, M., Hu, Z., Lu, X., et al.: Can LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs. arXiv preprint (2023), https://arxiv. org/abs/2306.13063, arXiv:2306.13063

  14. [22]

    Xu, F., et al.: Towards large reasoning models: A survey of reinforced reasoning with large language models (2025), https://arxiv.org/abs/2501.09686

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.