List of Papers Browse by Subject Areas Author List
Abstract
Medical Multimodal Large Language Models (MLLMs) hold significant potential for clinical visual question answering (VQA), yet their reliability is often compromised by unstable visual grounding and flawed reasoning logic. While reinforcement learning-based post-training, such as Group Relative Policy Optimization (GRPO), offers a path to steer model behavior, its efficacy is severely constrained by the high cost of expert supervision. To address this, we propose VRUGA, an active learning framework designed to maximize GRPO post-training efficiency under a constrained expert budget. Specifically, we introduce Dual-Uncertainty Quantification (DUQ) to identify high-risk failure modes by jointly evaluating visual grounding stability and reasoning trajectory consistency. Logic-Aware Diversity Sampling (LADS) is subsequently employed to select a representative active set by leveraging decoder hidden-state embeddings to capture a diverse spectrum of latent reasoning patterns. Finally, a Cross-Threshold Pseudo-Labeling (CTPL) mechanism stabilizes the alignment process through tiered supervision, integrating expert-corrected samples with majority-voted pseudo-label anchors to mitigate policy overfitting. Extensive evaluations on VQA-RAD, SLAKE, and PathVQA benchmarks demonstrate that VRUGA achieves a 75.2% average accuracy, significantly outperforming state-of-the-art active learning baselines while requiring only a 10% expert annotation budget.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3198_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
N/A
Link to the Dataset(s)
slake dataset: https://www.med-vqa.com/slake/
BibTex
@InProceedings{XiaBoy_VRUGA_MICCAI2026,
author = { Xiao, Boyi AND Wu, Jianghao AND Zhang, Rongzhao AND Xu, Jie AND Zhao, Feng AND Wang, Guotai AND Zhang, Shaoting},
title = { { VRUGA: Visual and Reasoning Uncertainty Guided Active Learning for Medical MLLM Post-training } },
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
This paper proposes VRUGA, an active learning framework for GRPO-based post-training of medical MLLMs under limited expert annotation budget. The framework has three modules: (1) DUQ measures visual attention consistency and CoT token entropy as dual uncertainty, (2) LADS performs diversity sampling in decoder hidden-state space using farthest point sampling, and (3) CTPL combines expert labels with pseudo-label anchors to stabilize RL post-training. Experiments on VQA-RAD, SLAKE, and PathVQA show 75.2% average accuracy with 10% expert budget.
- 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.Budget-efficient post-training of medical MLLMs is a practical bottleneck. The connection between active learning and RL alignment is clearly motivated.
2.Separating visual grounding instability (U_vis) from reasoning trajectory uncertainty (U_rsn) goes beyond simple logit entropy.
- 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.No standard deviations, confidence intervals, or significance tests are reported. As a result, the robustness of the reported gains cannot be confidently assessed from the current submission alone.
2.The strongest reported result, 75.2%, uses 10% expert data plus 50% unlabeled anchors, whereas the main baselines in Table 1 are reported under 10% expert data only. The paper does disclose this and also reports VRUGA (w/o Unl. ) at 73.0%, which is the fairer direct comparison. However, the headline gain should be framed more carefully. I would encourage the authors either to reframe the headline claim or to add a matched ‘baseline + pseudo-label anchor’ (e. g. , “DEUCE + CTPL”) comparison.
3.The paper motivates DUQ using visual grounding instability and reasoning trajectory uncertainty, but evaluation is limited to exact-match accuracy on closed-ended VQA. This is a standard metric for the benchmark, yet insufficient to directly validate claims about reasoning quality or grounding faithfulness.
- 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.Please clarify the exact GRPO reward function for both expert-labeled and pseudo-labeled samples. Is it exact match only, or does it include format/CoT quality components?
2.How is the train/test split constructed? Is the standard split from each benchmark used, or is there a custom split for the active learning pool?
3.What is the computational overhead of DUQ (K=8 stochastic runs with attention extraction) compared to standard baselines?
- 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 paper sits at a useful intersection of AL and GRPO post-training that has received little prior attention in the medical VQA literature. However, several issues prevent acceptance in the current form. First, the absence of statistical evidence makes the reported gains difficult to assess with confidence, particularly on VQA-RAD where the margin is very small. Second, the headline comparison involves asymmetric data usage between VRUGA and baselines. These are important issues that the authors would need to address before acceptance can be considered.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #2
- Please describe the contribution of the paper
The paper proposes VRUGA, an active learning framework for expert-efficient GRPO post-training of medical MLLMs for VQA. The method combines three components: DUQ, which scores samples using both visual grounding instability and reasoning uncertainty; LADS, which selects a diverse subset of high-risk cases in decoder hidden-state space; and CTPL, which stabilizes post-training by mixing expert-labeled high-uncertainty samples with pseudo-labeled low-uncertainty anchor samples. Experiments on VQA-RAD, SLAKE, and PathVQA show improved accuracy under a 10% expert annotation budget, with the full method reaching 75.2% average accuracy.
- 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 proposed question is well-motivated: how to use limited expert supervision effectively when post-training medical MLLMs. 2.The proposed method are coherent and reasonable. 3.I appreciate the comprehensive experiments.
- 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.When talking about AL, we care about the efficiency. DUQ requires K = 8 stochastic generations per sample, so i wonder how about the running time? Will it increase the computational burden? 2.No CI showed, so how to make sure it’s statistically significant?
- 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?
I like this work as it’s researching on a growing field, and the proposed method is reasonable. But i dislike the result they are showing.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Review #3
- Please describe the contribution of the paper
The paper proposes VRUGA, an active learning framework that enhances reinforcement-learning-based post-training (GRPO) for medical multimodal LLMs, making them more label-efficient for clinical VQA. It introduces Dual-Uncertainty Quantification (DUQ) by combining visual grounding instability and reasoning uncertainty, Logic-Aware Diversity Sampling (LADS) for reasoning trajectory sampling in the decoder hidden-state space, and Cross-Threshold Pseudo-Labeling (CTPL) to stabilize GRPO with expert-labeled hard cases and majority-voted pseudolabel anchors. The framework achieves state-of-the-art accuracy under a limited expert budget across multiple VQA datasets.
- 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 proposes a dual-view uncertainty framework that combines visual grounding instability with reasoning-trajectory entropy. This design is interesting as it explicitly models uncertainty in both visual grounding and reasoning processes for generative multimodal VQA.
2.The proposed LADS strategy leverages decoder hidden states of CoT tokens and applies farthest-point sampling weighted by the dual-uncertainty score. This provides a logic-aware way to capture diverse reasoning patterns during sample selection.
3.The CTPL module incorporates high-confidence unlabeled samples into GRPO via majority-voted pseudo-labels, thereby improving data efficiency and stabilizing training under limited labeled data.
- 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 method relies on attention maps as proxies for visual grounding stability (Uvis). However, the reliability of attention as an explanation mechanism is debated, and the paper does not clearly specify how attention is extracted (e.g., layers/heads, aggregation scheme, or the top-m token selection). 2.DUQ requires multiple stochastic decodes for each sample across the unlabeled pool, which could be computationally expensive for large datasets. The paper does not discuss runtime or scalability. 3.The framework assumes the availability and usefulness of Chain-of-Thought (CoT) reasoning in medical VQA, but the potential mismatch between generated CoT and the model’s actual reasoning process is not analyzed. 4.The main comparison (Table 1) uses additional unlabeled data for the proposed method, while the baselines appear to rely only on labeled data. Including semi-supervised or self-training counterparts for baselines would make the comparison fairer.
- 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
Q1.How exactly are attention maps extracted for Uvis (e.g., which layers/heads and aggregation strategy)? What is the value of m for top-m visual tokens?
Q2.Are the reported results averaged over multiple random seeds for DUQ sampling, LADS selection, and GRPO training? If not, could variance or error bars be reported?
Q3.What is the computational overhead of DUQ (K stochastic decodes per sample)? It would be helpful to report runtime or GPU cost.
Q4.Do Uvis and Ursn correlate with actual error likelihood? Correlation or calibration analysis would help validate the uncertainty formulation.
- 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 proposes an interesting framework that combines dual-uncertainty estimation with logic-aware diversity sampling and pseudo-labeling for data-efficient multimodal VQA. The idea is promising and addresses an important challenge in reducing expert annotation requirements for medical VQA tasks. While several aspects could be better analyzed—such as the reliance on attention maps for grounding estimation, the computational cost of DUQ, and the assumptions regarding Chain-of-Thought reasoning—these issues mainly relate to missing analysis or experimental validation rather than fundamental flaws in the approach. Additional experiments, clearer methodological details, and fairer comparisons with semi-supervised baselines would further strengthen the work. Overall, the proposed ideas are interesting and potentially useful, and with some clarifications and stronger empirical analysis, the paper could make a meaningful contribution.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
N/A
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
N/A
Meta-Review
Meta-review #1
- Your recommendation
Provisional Accept
- 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 MLLM post-training in an active learning setting where the visual uncertainty, reasoning uncertainty, and logic-based sampling are used to select data for GRPO optimization. The paper is of high quality with clear presentation, pretty illustration, and extensive experimentation. The results demonstrate improvement with limited expert annotation. Considering such practical value, I recommend acceptance.