List of Papers Browse by Subject Areas Author List
Abstract
Medical vision language models have demonstrated strong potential for medical image understanding and reasoning. However, their interpretability and reliability remain limited because medical vision language tasks require precise visual grounding and clinically coherent explanations. This is particularly challenging due to complex imaging patterns and scarce expert annotations. These challenges reduce the effectiveness of conventional reasoning methods and lead to visual hallucination, defined as the generation of text that is not grounded in the given visual input, which further limits their applicability in clinical practice. To mitigate these challenges, we propose an interpretable medical inference (I$^2$-Med) framework that enhances visual faithfulness during inference without additional training to improve reasoning and explanation faithfulness in medical vision language models. During text generation, I$^2$-Med uses a visual guidance module to check whether each candidate word is supported by the medical image and consistent with the current reasoning. Based on this guidance, the model’s output scores are adjusted to favor visually grounded words, without changing the model parameters. Extensive experiments on the OmniMedVQA benchmark, covering eight medical imaging modalities and five clinical question types, demonstrate that our proposed I$^2$-Med framework generates more clinically meaningful and visually faithful reasoning chains and reduces visual hallucination under standard medical VQA evaluation metrics.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/5659_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/mtanveer1/l2-med
Link to the Dataset(s)
OmniMedVQA dataset: https://openxlab.org.cn/datasets/GMAI/OmniMedVQA
BibTex
@InProceedings{SajMd_I²Med_MICCAI2026,
author = { Sajid, Md AND Maheshwari, Shreeyut AND Mishra, Akshat AND Mishra, Ritik AND Tanveer, Mohammad},
title = { { I²-Med: Interpretable Medical Inference Through Visual-Guided Dynamic Logits Calibration } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16882},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes an interpretable medical inference framework for medical VLMs that aims to improve reasoning, explanation faithfulness, and reduce hallucination. The framework applies a GRPO-based structured reasoning strategy, which encourages the medical VLM to output structurally consistent and diagnostic-accurate answers. To ensure medically grounded consistency between input images and generated text, an inference-time visual-guided logits regulation module is introduced, leveraging a pretrained vision-language model to reweight token probabilities based on estimated image-text alignment. The method is evaluated on the OmniMedVQA benchmark, covering eight medical imaging modalities and five clinical question types.
- 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.This paper addresses hallucination and interpretability in medical VLMs, which are critical challenges for reliable deployment in clinical settings. 2.The proposed combination of GRPO-based medical reasoning and logit calibration allows the method to be integrated with existing VLMs without substantial architectural changes.
- 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 paper claims improvements in medical reasoning and interpretability. However, the introduced structured policy optimization adopts a standard GRPO strategy without specific adaptations, which is a standard reinforcement learning strategy and is not specifically tailored to medical VLMs. 2.The calibration mechanism involves several heuristic components (e. g. , the formulations of relative clinical grounding and adaptive clinical constraint) without clear theoretical analysis. Visual grounding is only introduced via post-hoc logit calibration during inference, while the training objective does not incorporate any visual alignment signal.
- 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
Fig. 2 is difficult to interpret, the notation is inconsistent with the main text, the figure appears to focus primarily on the calibration module without clearly illustrating the optimization process.
- 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?
See major strengths and weaknesses.
- 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 authors solved most of my questions and will release their codes. I keep my positive rating.
Review #2
- Please describe the contribution of the paper
This paper proposes I²-Med, an interpretable medical inference framework designed to improve visual faithfulness and reduce hallucinations during reasoning. The method combines a GRPO-based structured reasoning stage with an inference-time visual-guided logits calibration mechanism that leverages BiomedCLIP to assess token-level image-text consistency, without requiring additional model training. Experiments on OmniMedVQA, covering eight imaging modalities and five medical VQA task types, demonstrate competitive performance.
- 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 addresses an important problem in medical VLMs, the lack of visual faithfulness and interpretability during reasoning. 2.The proposed method is technically appealing because it improves grounding at inference time without modifying model parameters, making it practical and lightweight. 3.The idea of using an external medical vision-language model to guide token decoding is interesting and potentially useful for improving the trustworthiness of medical VLM outputs.
- 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.Inference Overhead: It is unclear how many candidate tokens are evaluated at each decoding step. If visual consistency is computed for every candidate token, this may introduce substantial inference overhead. The paper should clearly report the additional runtime and computational overhead compared with standard inference. 2.Potential Noise from Token-Level Visual Alignment: Not every generated token is expected to align directly with the image. Applying visual alignment to every token may introduce unnecessary noise, especially for function words or reasoning-related tokens that are not expected to have direct visual correspondence. 3.Unclear Base Model and Training Details: The proposed model first uses GRPO-based post-training and then applies Dynamic Logits Calibration during inference. However, the exact base model and training details are not clearly specified. These factors can substantially affect the final performance. Without this information, it is hard to judge how much of the gain comes from the proposed calibration itself rather than from a strong base model or training recipe. 4.Lack of Ablation Study: The current submission lacks an ablation study, making it unclear how much each component contributes to the observed gains, including the proposed Visual-Guided Dynamic Logits Calibration, the first-stage post-training, and the inherent capability of the base model. 5.Limited Analysis of Reasoning and Grounding: Although I²-Med can correct the final answer, it remains unclear whether the reasoning process itself is also improved. Since the method emphasizes dynamic token-level visual grounding, the paper should provide qualitative or quantitative analysis showing which tokens are corrected and what image evidence supports these corrections. 6.Limited Evaluation Benchmark: The method is evaluated on only a single benchmark, which is somewhat limited.
- 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 provide sufficient information for 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.
(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?
1.Potential Noise from Token-Level Visual Alignment 2.Unclear Base Model and Training Details 3.Lack of Ablation Study
- 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.
The authors have addressed my concerns, and I recommend acceptance.
Review #3
- Please describe the contribution of the paper
The paper addresses the important problem of improving visual faithfulness and reducing hallucination in medical vision-language modeling. Its main contribution is the combination of structured reasoning (GRPO‑style) with inference‑time, visually‑guided logits calibration using BiomedCLIP‑based visual scoring, adapted specifically to the medical VLM setting. The work also provides a broad evaluation across multiple modalities and task types.
- 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.Addresses an important problem (improving visual faithfulness and reducing hallucination in medical VLM). 2.Interesting idea (combining structured reasoning with inference‑time visual‑guided logits calibration). 3.Broad evaluation across multiple modalities and task types.
- 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 framework mainly combines existing ingredients (GRPO reasoning, decoding-time calibration, BiomedCLIP scoring) with adaptation to the medical VLM setting as the main contribution.
2.Structured reasoning and visually-guided decoding do not by themselves establish strong interpretability; the paper lacks a clear discussion of what interpretability is actually provided and how it differs from simply producing more structured outputs.
3.Most results rely on standard VQA accuracy, which does not fully validate the central claims about visual grounding and hallucination reduction; stronger faithfulness-oriented evaluations are needed.
4.The paper would benefit from more targeted ablations to isolate the contributions of the structured reasoning stage vs. visual calibration stage, and to show sensitivity to design choices such as top‑k candidate selection and the BiomedCLIP-based calibration scheme.
- 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 provide sufficient information for 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 in medical vision-language modeling, namely improving visual faithfulness and reducing hallucination during medical inference. The idea of combining structured reasoning with inference-time visual-guided logits calibration is interesting, and the broad evaluation across multiple modalities and task types is a positive aspect. However, I find the overall contribution somewhat borderline, and my final inclination is Weak Accept . 1.The paper addresses a meaningful problem, but the methodological novelty appears somewhat limited. The overall framework mainly combines existing ingredients, including GRPO-style structured reasoning, decoding-time logits calibration, and BiomedCLIP-based visual scoring, with the main contribution being their adaptation to the medical VLM setting. 2.The interpretability claim is not fully convincing. While the method encourages structured reasoning and visually guided decoding, this does not by itself establish strong interpretability. The paper would benefit from a clearer discussion of what kind of interpretability is actually provided and how it should be distinguished from simply producing more structured outputs. 3.The experimental results are reasonably broad, but the evidence for reduced hallucination and improved faithfulness is still somewhat indirect. Most results are based on standard VQA accuracy-style evaluations, which do not fully validate the central claims about visual grounding and hallucination reduction. More stronger faithfulness-oriented evaluations would make the paper more convincing. 4.The gains over strong baselines are not entirely decisive. While the proposed method outperforms several zero-shot and constrained baselines, it does not clearly surpass the strongest unrestricted reasoning baseline. This makes the practical advantage less conclusive. 5.The paper could be strengthened by more targeted ablations and analysis. It would be helpful to better isolate the contributions of the structured reasoning stage and the visual-guided calibration stage, and to show how sensitive the method is to design choices such as top-k candidate selection and the BiomedCLIP-based calibration scheme.
- 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 rebuttal does not fully resolve all concerns (e.g., limited novelty, weak faithfulness evaluation), but it provides reasonable clarifications: the contribution is positioned as a lightweight inference‑time calibration framework, interpretability is explicitly defined as consistency between image, reasoning, and prediction, and code with sensitivity analyses will be released. Given the practical importance of reducing hallucinations in medical VLM, the paper is acceptable.
Author Feedback
R1-5.1/R3-5.1: We thank the reviewer for the observation. Our main contribution is the proposed inference-time Dynamic Logits Calibration framework built on existing medically aligned reasoning backbone, rather than new GRPO-based optimization method. The framework is lightweight & adaptable, improving reasoning consistency across medical VLMs without additional backbone retraining.
R1-5.2.We agree that the proposed calibration includes heuristic design choices; however, each component targets a specific aspect of inference-time grounding. Relative clinical grounding strengthens consistency between generated reasoning & medical image evidence, while the adaptive clinical constraint suppresses visually inconsistent token generations during decoding. Our main objective is to improve reasoning consistency entirely at inference time without modifying backbone parameters.
R1,R2,R3-7: To preserve author anonymity, we did not include a public code link. We are determined to release source code and inference pipeline upon acceptance.
R1-9: Thank you for pointing out the notation inconsistency in Fig 2.We revised the figure & will add it to final paper.
R2-5.1: Visual consistency is computed only for the top-30 candidate tokens at each decoding step, introducing limited inference overhead. Runtime analysis and top-k sensitivity studies will be added to the final version(see R*).
R2-5.2: Our method performs calibration over contextual sliding windows rather than independent tokens, preserving semantic consistency & reducing noise effect. Window-size sensitivity (8,12 & 16) reproducibility will be given in source code(see R*).
R2-5.3: We acknowledge that the manuscript wording may have created ambiguity regarding additional GRPO-based training(See R1-Q7.1).
R2-5.4: See R* & response to R2-Q7-1,2; we will clarify this further in the final version & source code.
R2-5.5: The reasoning backbone may generate intermediate reasoning that is not fully grounded in the medical image. Our calibration framework improves consistency between reasoning trajectory, image evidence & final prediction, supported quantitatively in Tabs. 1-2 & qualitatively in Fig. 3.Additional qualitative analyses & visualizations will be released with the public repository (See R*).
R2-5.6: We agree that evaluation on additional benchmarks would strengthen the paper. However, we selected OmniMedVQA due to its diversity across 8 imaging modalities & 5 clinically relevant task categories, making it suitable to evaluate reasoning consistency & cross-modality generalization in a unified setting. Due to MICCAI rebuttal guidelines, we are unable to add new experiments at this stage & will mention extended benchmarking as future work.
R3-5.2: We agree that structured outputs alone do not establish interpretability. Existing reasoning-based medical VLMs may generate plausible but weakly grounded reasoning inconsistent with image evidence. In our work, interpretability refers to improving consistency between medical image, reasoning trajectory & final prediction, as shown in Fig 3. R3-5.3: While VQA accuracy is the primary quantitative metric, our goal is also to reduce visually inconsistent reasoning generations. We therefore included qualitative analyses showing improved grounding consistency & fewer reasoning failures [Fig 3]. (Further, see R*; R2-Q7-1,2,4)
R3-5.4: Thanks for your suggestion, please see remark R* and responses to R2-Q5-1,2,4
R3-10.4: We agree that gains over the strongest unrestricted reasoning baseline are modest in some settings. However, the proposed framework consistently improves reasoning grounding & prediction consistency across modalities while operating purely at inference time without modifying backbone parameters.
R3-10.5: Thanks for your suggestion, please see R*, & responses to R2-Q5-1,2,4; R3-Q5-4
Remark (R*): These results were originally included in the supplementary material before its removal to comply with MICCAI guidelines.
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 reviewers see the paper as a practical and well-motivated attempt to tame hallucination. But major concerns were raised: the method mostly recombines existing ideas, relies on heuristic calibration without strong theory, lacks ablations and clear attribution of gains, and provides limited evidence that it truly improves reasoning faithfulness rather than just final answers. The authors are encouraged to address them in the rebuttal.
- 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 paper improves visual faithfulness and reducing hallucination during medical reasoning.
The rebuttal reasonably clarifies the method design, ablation results, inference overhead, and planned code release.
Given the practical value of a lightweight inference-time calibration framework and the reviewers’ generally positive post-rebuttal opinions, I recommend acceptance, while asking the authors to clearly discuss the remaining limitations on novelty, interpretability, and faithfulness evaluation.
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.
Reviewers have acknowledged that most raised concerns, including clarity, technical contributions, noise in alignment, ablation studies, and limitations/overclaims, are clarified in a reasonable manner. This leads to consistently positive ratings after the rebuttal.
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.
All reviewers highlight the contributions of this submission and would like to accept.