List of Papers Browse by Subject Areas Author List
Abstract
Multimodal Large Language Models (MLLMs) offer transformative potential for medical diagnostics, yet their tendency to generate plausible but ungrounded hallucinations remains a critical barrier. In black-box settings, where internal model states are inaccessible, existing uncertainty quantification methods like visual-induced agreement rely on final label consistency, which often fails to capture systematic overconfidence. In this work, we move from visual stability to reasoning stability, identifying a failure mode where models maintain consistent labels despite divergent and fabricated logic. We introduce Reasoning Trace Divergence (RTD), a novel empirical signal that exposes these latent hallucinations by probing the stability of the model’s logical paths under stochastic visual perturbations. Evaluating this signal on two patch-level histopathology datasets using Gemini and MedGemma, we demonstrate that reasoning instability is a robust predictor of grounding failures. By filtering for high reasoning stability, we identify a safe operating zone that significantly improves classification reliability, providing a principled direction for trustworthy black-box medical AI. The code is available at https://github.com/anastlibr99/reasoning-trace-divergence-blackbox-mllms-histopath
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/6450_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/anastlibr99/reasoning-trace-divergence-blackbox-mllms-histopath
Link to the Dataset(s)
NCT-CRC-HE-100K: https://huggingface.co/datasets/DykeF/NCTCRCHE100K
PatchCamelyon: https://github.com/basveeling/pcam
BibTex
@InProceedings{KazAna_Reasoning_MICCAI2026,
author = { Kazmina, Anastasiia AND Venhoff, Constantin AND Wong, Bryan AND Hong, Sungrae AND Xie, Yutong AND Yi, Mun Yong},
title = { { Reasoning Trace Divergence: An Empirical Signal for Trustworthy Black-Box MLLMs in Histopathology Classification } },
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
The paper identifies a potential failure mode in multimodal large language models, namely systematic overconfidence under few-shot in-context learning, where predictions remain stable while reasoning varies. It proposes a novel uncertainty estimation metric, Reasoning Trace Divergence (RTD), which measures the stability of model-generated reasoning by converting free-form explanations into structured feature representations and computing entropy under semantic-preserving perturbations.
- 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 uncertainty estimation for black-box MLLMs in high-stakes medical settings, which is timely and clinically relevant. 2.The transition from label-level consistency to reasoning-level consistency is intuitive and well-articulated. 3.The paper is generally well-written, with a structured formulation and clearly defined pipeline.
- 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 feature space is automatically extracted by an LLM from generated reasoning, without grounding in established clinical criteria. There is no guarantee that the extracted features correspond to true diagnostic evidence used by clinicians. Consequently, the derived entropy over these features may not reflect clinically relevant uncertainty. 2.Insufficient discussion of ablation results. Table 3 shows that Visual Stability dominates AUROC improvement while Reasoning Stability dominates Top-10% accuracy. The authors only note that combining both is best, without discussing this complementary pattern. 3.The proposed framework does not explicitly exploit characteristics unique to histopathology images. Therefore, it is unclear why the study is limited to this modality, rather than including other medical imaging modalities. 4.The method relies on a secondary MLLM (Gemini 2.5 Flash Lite) to analyze reasoning traces, introducing another black-box model whose own uncertainty is not accounted for.
- 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.
(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) the LLM-extracted features lack clinical grounding; (2) the ablation study’s complementary pattern (Visual for AUROC, Reasoning for top-10% accuracy) is noted but not discussed; (3) the framework does not exploit histopathology-specific characteristics; (4) using a secondary MLLM introduces unaccounted black-box uncertainty.
- 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 have addressed all my concerns. I recommend acceptance.
Review #2
- Please describe the contribution of the paper
This paper proposes a metric to quantify instability in vision language model outputs. This metric takes into account the outcomes of semantic-preserving perturbations to the images, and quantifies via entropy changes in features determined by text justifications. The metric is implemented in two datasets dealing with the problem of patch classification in histopathology, and compared with other hallucination detection methods.
- 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 problem of reliability in complex, black-box model is relevant and challenging as more modalities are considered.
- The definition of the metric for instability takes into account simple concepts (domain specific) and aggregates them to provide a notion of stability.
- 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.
-
Unclear aspects in the methodology make it hard to follow the approach and appreciate how it was built up: 1.Generation of justifications along with categorical predictions: Does the system prompt includes an instruction to generate the language reasoning? 2.Another meta-analytical prompt is introduced without explanation. How are those unique morphological features defined? 3.How is visual entropy H_vis defined? not explained before, briefly mentioned in section 2.2.Related to this point, in the ablation test, how was the visual and reasoning stability separated? 4.Explain how the definition of the safety operating zone can be used during inference time with new samples. Currently, it seems to be a post-hoc analysis.
-
Concerns with claims and experimental design choices: 1.The use of different E values for the zero-shot vs few-shot comparison but no comments about these results or discussion. 2.Clipping entropy values between 0 and 1 may not result in a calibrated confidence. 3.Only a subset of the evaluation sets in the datasets was used for the experiments and artificially balanced. From my experience with histopathology data, one of the challenges is having a single label for a high resolution image, where individual patches may not have informative features for that category, so it is a highly imbalanced class. 4.The work aims to compare a general-purpose foundation model with a medically-tuned model, but for Gemini was used for the reasoning traces. It was unclear whether this will be an issue, given that I did not understand how those morphological features (which have clinical relevance) are defined. 5.Missing justification for why MedGemma “refused” to answer when the harder transformations where applied, which I also wonder if they are clinically feasible. 6.No clear support (previous literature) for why the diagnostic knowledge remains invariant regardless of spatial transformation.
-
- 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
- In the method explanation, refer to the overview in figure 1.
- In figure 2, I would suggest also showing some examples of the actual text in the reasoning section to show the added value of providing these.
- Histopathology is first mentioned on page 2.I suggest to present it as an application that can benefit from MLLMs in the first paragraph.
- Review proper use of acronyms, e.g., ICL, VLM, D4.
- May need to explain concepts like gland architecture or serration patterns for readers without clinical background.
- In 3.2., there are only 3 strategies listed, fix the (4).
- 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?
There are many unclear points in the methodology listed above and choices not well justified. The paper starts with some motivations, such as the idea of comparing general-purpose LLMs and domain-specific, and zero/few-shot prompting, but the execution of the study questions if these goals are properly addressed.
- Reviewer confidence
Somewhat confident (2)
- [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.
The rebuttal includes additional descriptions that clarify some doubts I had during the review process, e.g., regarding model selection, computations, and ablation. However, I still identify a few concerns about the justification of the approach (both conceptually and empirical evaluation), and as mentioned by another reviewer, the reliability of this uncertainty quantification. In the response, it is mentioned that the concepts are aligned with established medical descriptors, but did not explain how the alignment is determined (reviewed by experts?).
Review #3
- Please describe the contribution of the paper
The paper addresses the important problem of trustworthiness and uncertainty estimation in multimodal large language models (MLLMs), specifically in the context of histopathology classification. A key challenge in such models is their tendency to produce overconfident predictions despite hallucinations or incorrect reasoning.
To tackle this, the authors propose Reasoning Trace Divergence, a framework that quantifies uncertainty by combining two complementary signals: (1) visual stability, measured via entropy over predictions obtained from perturbed inputs, and (2) reasoning consistency, derived from analyzing textual explanations generated by the MLLM.
The latter is enabled by leveraging an auxiliary LLM to categorize and structure the generated reasoning traces, allowing the computation of an entropy-based divergence measure over explanations. These two signals are combined into a unified uncertainty estimate, which is then used to recalibrate prediction confidence, lowering confidence for incorrect predictions and increasing it for correct ones.
The method is evaluated on two histopathology datasets and compared against state-of-the-art uncertainty estimation approaches, demonstrating strong empirical improvements.
- 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 tackles a highly relevant and timely problem, as reliability and calibration of MLLMs are critical barriers to their deployment in clinical practice. The motivation is clearly articulated and extends beyond histopathology to many multimodal medical applications.
A key strength is the novel combination of visual perturbation-based uncertainty with reasoning-based analysis. In particular, leveraging reasoning traces as an uncertainty signal is an interesting and original idea that goes beyond standard entropy-based or ensemble-based approaches.
The approach is conceptually elegant and well structured, separating visual and textual sources of uncertainty while combining them in a meaningful way.
The experimental results are convincing and mostly comprehensive, showing clear improvements over strong baselines across two datasets. The ability to improve confidence calibration (increasing confidence for correct predictions and decreasing it for incorrect ones) is particularly compelling.
Overall, the paper provides a practically relevant contribution toward more trustworthy MLLM-based systems.
- 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.
While the results are strong, the evaluation of confidence calibration could be more comprehensive. The analysis focuses on a “safe operating zone” defined as the top 10% most confident predictions. While this is informative, it would be more convincing to include performance curves across different confidence thresholds (e.g., top 15%, 20%, etc.), which would provide a more complete picture of the method’s behavior.
Related to this, the paper does not explicitly analyze the correlation between predicted confidence and actual accuracy. Standard calibration metrics or correlation analyses would strengthen the claims regarding improved reliability.
On a minor note, there appears to be a visual inconsistency in Figure 1, where the example image for the NORM class (lower half, top row) should match the corresponding query image in the upper half.
- 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.
(5) Accept — should be accepted, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
This paper presents a novel and well-motivated approach to uncertainty estimation in MLLMs, addressing a critical limitation of current systems—namely, overconfidence in the presence of hallucinations.
The proposed idea of combining prediction entropy under visual perturbations with reasoning trace divergence is both intuitive and innovative, and the empirical results convincingly demonstrate its effectiveness.
While the evaluation could be further strengthened with more detailed calibration analysis, these are relatively minor issues that do not detract from the overall contribution.
- 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.
I rated this paper as accepted in my initial review, independent of the rebuttal. The paper presents a novel and well-motivated approach to uncertainty estimation in MLLMs, combining visual perturbation entropy with reasoning trace divergence in a conceptually elegant way, and demonstrates convincing empirical improvements across two histopathology datasets. My initial concerns were minor: a more comprehensive calibration analysis across multiple confidence thresholds and a correlation analysis between predicted confidence and accuracy. These points are not directly addressed in the rebuttal, but they do not detract from the overall contribution and can reasonably be addressed in the camera-ready version. The visual inconsistency in Figure 1 is acknowledged and will be corrected, and code release is committed upon acceptance. I therefore maintain my recommendation to accept.
Author Feedback
We thank all reviewers for their feedback. We are encouraged by the shared recognition of the problem’s clinical relevance and address all concerns below.
Q1: Feature Extraction and Clinical Grounding (R1-1, R2-1-2, R2-2-4) In response to concerns regarding how morphological features are defined and their clinical grounding, the meta-prompt jointly analyzes all 8 D4 reasoning traces, identifies recurring morphological concepts, and merges semantically equivalent expressions under a unified term, without relying on a predefined ontology. The resulting concepts align with established histopathological descriptors such as glandular architecture and cellular atypia, supporting their interpretability. We acknowledge the concern regarding explicit clinical grounding. This will be more precisely addressed in the Methodology section and in Future Work.
Q2: Secondary LLM and Model Comparison (R1-4, R2-2-4, R2-2-5) The secondary LLM performs constrained structured extraction rather than diagnosis generation; any variability makes results a conservative lower bound. Gemini is used only as a structured parser, preserving the integrity of the general-purpose vs domain-specific comparison. MedGemma frequently abstained under severe perturbations due to safety-oriented filtering; VASE was therefore omitted for fair comparison.
Q3: Ablation and Calibration (R1-2, R2-2-2, R3) In response to reviewers’ questions on the ablation and complementary signals, visual stability captures label inconsistency across D4 perturbations (driving AUROC), while reasoning stability captures consistency between predictions and structured explanations (improving Top-10% accuracy). Both are computed from the same D4 perturbations: Hvis from prediction consistency and reasoning stability from semantic feature consistency in reasoning traces. For the ablation, we use T=1 sampling as baseline, replace sampling noise with D4 perturbations, then add reasoning-trace stability under the same perturbations before combining both signals. Clipping only normalizes entropy to [0,1] before fusion, not calibration. We will clarify this setup further in the final version. RTD is an empirical signal whose complementary behavior motivates future theoretical study.
Q4: Inference-Time Applicability (R2-1-1, R2-1-4) Our framework requires no ground truth at inference time. The system prompt explicitly instructs the model to generate both a class prediction and structured reasoning in JSON format. A meta-prompt aggregates the 8 reasoning traces into a morphological feature frequency table, from which CUVR is computed in a fully unsupervised manner. The resulting confidence threshold can be applied directly to unseen samples.
Q5: Histopathology Specificity and D4 Justification (R1-3, R2-1-3, R2-2-6) Histopathology was selected because tissue orientation is not diagnostically meaningful, making D4 transformations provably semantic-preserving by a fundamental biological property. Hvis is the Shannon entropy of predicted labels across these 8 transformations; its formal definition will be added alongside Eq. 3.We evaluated on two datasets with different tissue types and tasks to validate within-modality generalizability. The framework naturally extends to other modalities where equivalent perturbations can be clinically justified.
Q6: Dataset Balancing and E Values (R2-2-1, R2-2-6) Regarding dataset construction and few-shot settings, PCam and NCT-CRC-100K provide independent patch-level labels, making each patch a valid evaluation unit with balanced sampling across classes. We evaluate E∈{0,3,6} to study in-context learning effects on stability; while few-shot prompting may improve accuracy, it can increase reasoning coherence without reducing overconfidence, which is the failure mode targeted by RTD. We will further elaborate on this in the final version.
Figure 1 will be corrected and all R2 presentation points addressed. Code will be released upon acceptance.
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 request rebuttal on methodological clarity and validity, especially how reasoning traces, LLM-extracted morphological features, visual entropy, reasoning stability, and the “safe operating zone” are defined and used at inference time.
The authors should address concerns that the feature space and uncertainty estimates rely on a secondary black-box MLLM without clinical validation, and clarify whether these features correspond to established histopathological diagnostic criteria.
The rebuttal should also strengthen calibration and robustness evidence by adding threshold-performance curves, confidence–accuracy correlation or standard calibration metrics, deeper ablation discussion, justification of spatial transformations and dataset balancing, and correction of the Figure 1 inconsistency.
- 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 is accepted as a borderline but valuable contribution. Two reviewers support acceptance after rebuttal, including one who changed from weak reject to accept after the authors addressed concerns, while one reviewer maintains rejection due to remaining questions about conceptual justification, clinical grounding of extracted concepts, and uncertainty reliability. Overall, the paper presents a novel and well-motivated approach to MLLM uncertainty estimation by combining visual perturbation entropy with reasoning trace divergence, with convincing empirical results on histopathology datasets. Remaining issues, including clearer expert validation of concept alignment, broader calibration analysis, and minor figure corrections, should be addressed in the camera-ready version.
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.
The reviewers appreciate the significance of the problem and extensive experimental results, and the initial concerns have been mostly addressed during the rebuttal. I recommend accepting this submission.