acceptodds
Under review as a conference paper at ICLR 2027

A Training-Free Trustworthy Framework for Medical Image Diagnosis with Large Language Models

Abstract

Large language models (LLMs) enable training-free and concept-level interpretable medical image diagnosis. However, two issues prevent their clinical use: hallucination, under which erroneous predictions carry nearly the same high confidence as correct ones and caused the model untrustworthy, and the long-tailed class distribution of medical data, which makes common diseases overwhelm rare ones. To address these issues, in this paper, we propose a novel trustworthy diagnosis framework built entirely on frozen LLMs without any training. To achieve the trustworthy diagnosis, we design two inference pathways: a concept matching pathway that diagnoses based on medical concepts, and a visual perception pathway that diagnoses from visual features. Both pathways are further corrected by a long-tailed prior debiasing module to tackle the long-tailed distribution issue. Then, a Dirichlet evidence fusion module finally converts the two debiased posteriors into evidential concentrations, producing a trustworthy fused diagnosis together with a sample-wise uncertainty. Experiments on benchmark datasets show its effectiveness. Besides, the uncertainty of our method is aligned with the accuracy of the prediction, enabling hallucination detection that traditional confidence alone cannot provide.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
classification, medical image classification, LLM

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qbTUOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a training-free framework for medical image diagnosis built on frozen multimodal large language models. It first retrieves similar training images with a visual encoder to build in-context examples. It then diagnoses through two pathways: one predicts clinical concepts and infers the class from them, and the other predicts the class directly from the image. Both pathways remove the class-frequency prior to address long-tailed data, and their outputs are fused with Dirichlet evidence and Dempster–Shafer theory, which also yields an uncertainty score. Experiments on dermoscopy, chest X-ray, and fundus datasets with three backbones report the best results on two datasets and show that uncertainty is associated with prediction correctness.

Critical strengths

1 Trustworthy medical AI is an important research direction, especially when large language models may produce confident but incorrect diagnoses. The paper addresses this challenge without requiring expensive model fine-tuning, which is practically attractive.

2 The proposed dual-pathway design is reasonable. The concept pathway provides interpretable concept-level reasoning, while the visual pathway preserves direct visual information. Combining these complementary signals through evidence fusion is a meaningful design choice.

Critical weaknesses

1 The central claims of trustworthiness and hallucination detection are not supported. By construction, the uncertainty of the visual pathway depends only on the gap between the largest logit and the remaining logits. That gap is itself a conventional confidence score, which contradicts the claim that conventional confidence cannot separate correct from wrong predictions. The evidence consists of a modest difference in mean uncertainty between correct and wrong predictions, with no standard error-detection or selective-prediction metrics and no comparison against simple confidence baselines.

2 Novelty is limited, and the performance claims are not robust. The method mainly combines retrieval-based in-context learning, logit adjustment, and an existing evidential fusion scheme. On DDR, the dataset that best reflects the long-tail problem, CBVLM with Idefics3 achieves a higher Macro-F1 and a comparable BACC, which the text does not acknowledge. With test sets of a few hundred images, a single split, and no variance reported, gains of a few points are hard to confirm. A simple baseline that votes over the labels of the retrieved neighbors is also missing.

Questions for authors

1 How is the "traditional confidence" in Figures 1 and 4 computed

2 On DDR, why does the text omit that CBVLM with Idefics3 reaches a higher Macro-F1, and can you report per-class results for the minority grades and variance across splits?

3 Can the authors provide stronger evidence that the proposed uncertainty specifically detects hallucination rather than simply correlating with prediction difficulty?

4 Can the authors provide qualitative case studies showing the full reasoning pipeline, including retrieved examples, predicted concepts, pathway outputs, evidence fusion, and final uncertainty?

3 comments
AuthorsOpenReview ↗

Response to Reviewer qbTU [1/3]

[1/3] Due to the character limit, this response is split into three parts. This is part 1/3.

Thank you for your helpful suggestions and comments. We respond to the comments point by point in order below.

Weakness 1:

The visual logit gap can be used as a confidence score and determines the evidence strength of the visual pathway. To assess the reliability of the final diagnosis, we use the fused uncertainty, uFusedu_\mathrm{Fused}, which also depends on the concept evidence after conflict-based downweighting and on disagreement between the two pathways. Thus, two samples with the same visual logit gap can receive different fused uncertainty scores.

To evaluate the reliability of uFusedu_\mathrm{Fused}, we compare it with Confidence for detecting incorrect diagnoses on DDR using Qwen3-VL 8B. Both scores are evaluated against the same final diagnostic predictions and ground-truth labels. We sort samples independently by increasing uFusedu_\mathrm{Fused} and decreasing Confidence, then divide each ranking into five approximately equal-sized, non-overlapping quintiles. The first quintile contains the predictions ranked as most reliable by each score. We report the diagnostic error rate within each quintile. In the first quintile, the diagnostic error rate is 2.47% when ranked by uFusedu_\mathrm{Fused} and 17.28% when ranked by Confidence.

QuintileError rate by uFusedu_\mathrm{Fused} (%)Error rate by Confidence (%)
1st2.4717.28
2nd4.9419.75
3rd12.3516.05
4th24.6920.99
5th46.2516.25
AuthorsOpenReview ↗

Response to Reviewer qbTU [2/3]

[2/3] Due to the character limit, this response is split into three parts. This is part 2/3.

Weakness 2 and Question 2:

1. Methodological contribution

Our framework addresses two challenges in medical image diagnosis with frozen LVLMs: hallucination and bias toward common classes under long-tailed class distributions.

To mitigate hallucination, we design two complementary pathways because incorrect clinical findings predicted by a frozen LVLM can mislead the diagnosis. The concept matching pathway infers a diagnosis from predicted clinical findings and class-conditional concept statistics, while the visual perception pathway predicts the diagnosis directly from the image. Without training a separate evidence estimator, we use debiased class probabilities to allocate evidence across classes and sample-dependent logit gaps to determine the initial strength of that evidence. Because incorrect concept predictions can also have large logit gaps, we downweight concept evidence before fusion when the predicted findings conflict with co-occurrence patterns estimated from annotated data. The fused uncertainty reflects both pathways’ evidence strengths and the disagreement between their diagnostic predictions.

To mitigate long-tail bias, we introduce a separate prior correction strength for each pathway because class bias in a frozen LVLM’s outputs may differ from the bias implied by the empirical class frequencies of the target dataset. The concept-matching pathway adjusts the explicit class prior in its Bayesian diagnosis, while the visual-perception pathway adjusts the class probabilities derived from the frozen LVLM’s logits. This lets us control the degree of prior correction separately for each pathway.

Our experiments and ablation studies support the effectiveness of these designs in improving diagnostic performance and the reliability of uncertainty estimates.

2. Comparison of per-class ACC

We will revise the DDR discussion to explicitly report that CBVLM with Idefics3 achieves a higher Macro-F1 than our method with CheXagent, while the two configurations have comparable BACC. These are the backbones with which each method achieves its highest Macro-F1 on DDR. To explain this difference and assess performance on the minority classes, we report per-class ACC and F1 for both methods using Idefics3 on the same DDR test split.

SamplesOursCBVLM
ACC (%)F1 (%)ACC (%)F1 (%)
No DR21973.0678.2498.1796.85
Mild3063.3352.0553.3350.79
Moderate12053.3353.7879.1780.17
Severe1464.2951.4342.8646.15
Proliferative2166.6752.8347.6252.63

Our method achieves higher ACC and F1 on all three minority classes: Mild, Severe, and Proliferative. The F1 losses on No DR and Moderate outweigh the gains on the three minority grades, explaining the lower overall Macro-F1.

3. Variability across data splits

To assess how sensitive our method is to the choice of data split, we evaluated it with CheXagent on five DDR splits. The first is the official split, and the other four follow the same training, validation, and test set proportions, and the proportion of the sample sizes in each category across the three sets is also consistent. We report the results for each split and the mean and sample standard deviation across the five splits. These results quantify the method’s sensitivity to the choice of data split.

SplitBACC (%)Macro-F1 (%)
Split 164.3661.63
Split 262.8861.98
Split 370.6564.43
Split 471.0665.16
Split 567.7963.39
Mean ± Std67.35 ± 3.6763.32 ± 1.52

4. Retrieved-neighbor voting baseline

We evaluated a KNN (K=4K = 4) baseline on the original splits of all three datasets. We retrieve the four closest training images using frozen MedImageInsight features, then predict the label with the most votes. In a tie, we select the label of the closest neighbor among the tied classes. The BACC and Macro-F1 scores achieved on the three datasets are shown in the table below.

MethodDerm7ptCORDADDR
BACC (%)Macro-F1 (%)BACC (%)Macro-F1 (%)BACC (%)Macro-F1 (%)
KNN (K=4K = 4)82.6083.4567.1473.8552.2659.64
Ours-best88.0985.1378.7078.4364.3661.63
AuthorsOpenReview ↗

Response to Reviewer qbTU [3/3]

[3/3] Due to the character limit, this response is split into three parts. This is part 3/3.

Question 1:

In Figures 1 and 4 of the original submission (now Figure 6), “traditional confidence” is the maximum class probability of the final fused distribution, computed from the Dirichlet mean in Equation 12. This is the probability assigned to the final predicted class. The fused uncertainty uFusedu_\mathrm{Fused} reflects the total evidence supporting the diagnosis, so predictions with the same confidence can have different uncertainty values.

Conventional softmax confidence is typically computed from a classifier’s output logits, such as those of our visual pathway. For a fair comparison on the same final predictions, we use the maximum class probability after Dirichlet evidence fusion as the confidence baseline in our experiments.

Question 2:

See the response of weakness 2.

Question 3:

In our setting, hallucination can manifest as a prediction about an imaging finding that disagrees with the reference annotation. To examine how fused uncertainty responds to such errors, we compare the predicted clinical concepts with the reference annotations for each image. Figure 5 in Appendix L shows six selected cases from Derm7pt, CORDA, and DDR. In each case, the final diagnosis was correct, at least half of the clinical concepts were predicted incorrectly, and uFusedu_\mathrm{Fused} exceeded the corresponding test-set mean.

For example, in the bottom-left DDR panel in Figure 5, the concept pathway predicted hemorrhages, hard exudates, and microaneurysms, all of which were annotated as absent (red boxes 1–3). The fused prediction correctly classified the image as No DR, with uFused=0.3932u_\mathrm{Fused} = 0.3932, above the DDR test-set mean of 0.2338. In this case, the hallucinated findings contributed to elevated fused uncertainty even though the final diagnosis was correct.

Question 4:

We will include three qualitative cases showing the full diagnostic pipeline in Appendix F. In one CORDA case (Case 2 in Figure 3) evaluated with CheXagent, all four retrieved reference images were labeled COVID-19. The concept matching pathway predicted enlarged cardiomediastinum as present and the other seven findings as absent: cardiomegaly, lung opacity, edema, consolidation, pneumonia, pneumothorax, and pleural effusion. Based on these concept predictions, the concept matching pathway assigned a COVID-19 probability of 36.1% and predicted no COVID-19, while the visual perception pathway assigned a probability of 71.8% and predicted COVID-19. The visual pathway provided stronger evidence, and the conflict between the two pathways was 0.1270. After fusion, the COVID-19 probability was 61.66% and the final uncertainty uFusedu_\mathrm{Fused} was 0.2251. The fused diagnosis was COVID-19, matching the reference label.

Reviewer_hEnhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training-free framework for medical image diagnosis with frozen vision–language models, aiming to address high-confidence errors and class imbalance. It retrieves annotated reference images and combines two diagnostic paths: one based on predicted medical concepts and another based on direct visual assessment. Both paths apply class-prior adjustment, and their predictions are converted into evidence using logit gaps. Concept conflicts further reduce the evidence before fusion, producing a final diagnosis and a sample-level uncertainty score. Experiments cover three medical imaging datasets and three backbones, showing classification improvements on Derm7pt and CORDA and competitive results on DDR. Ablation studies examine the contributions of the main components.

Critical strengths
  1. The paper addresses clinically important problems: high-confidence diagnostic errors and class imbalance. The combination of concept-based reasoning and direct visual assessment is reasonable, as the two paths can provide complementary information. Adapting evidence fusion to frozen VLMs without additional fine-tuning is a useful contribution.

  2. The experiments cover three medical imaging datasets and three backbones, with good classification results on Derm7pt and CORDA. The ablation studies also support the contributions of the two diagnostic paths, prior adjustment, and fusion design to classification performance.

Critical weaknesses
  1. The trustworthiness claim is not sufficiently validated. Higher uncertainty on incorrect predictions and improved accuracy after filtering may also occur with standard confidence measures. Systematic quantitative comparisons with standard confidence measures on the same final predictions are needed to establish whether the proposed uncertainty score improves error detection. Moreover, the ablations report only classification metrics, leaving the uncertainty benefits of evidence construction and conflict-based downweighting unclear. These gaps limit the empirical support for the paper’s central claim of more trustworthy diagnosis. Since several components build on established ideas, the methodological contribution appears mainly incremental, making it particularly important to provide stronger evidence for the benefits of the paper’s specific designs.

  2. The method’s dependence on retrieved references is not sufficiently examined. Both diagnostic paths use an in-domain reference set with disease labels and concept annotations. However, the effects of reference-set size and retrieval quality are not evaluated. Only T=4 is reported, leaving unclear how sensitive the classification results and uncertainty estimates are to the number of retrieved examples. It is also unclear how the method behaves when the retrieved images are less relevant or dominated by one class, and whether its uncertainty score can flag errors in these cases. These missing analyses limit confidence in the method’s reliability when suitable annotated references are scarce or retrieval quality varies.

Questions for authors
  1. Does the proposed uncertainty score improve error detection over standard confidence measures on the same final predictions?

  2. Could you clarify how the baselines were adapted to the shared encoder and LLM setup, including whether key mechanisms such as BiomedCoOp’s BiomedCLIP-based prompt learning were preserved?

  3. How does changing the number or quality of retrieved references affect classification performance and the reliability of uncertainty estimates?

  4. Which model does “Qwen3-VL 7B” refer to? It seems that the official releases do not appear to include this 7B version, this inconsistency reduces my confidence in the precision of the reported experimental setup.

4 comments
AuthorsOpenReview ↗

Response to Reviewer hEnh [1/3]

[1/3] Due to the character limit, this response is split into two parts. This is part 1/3.

Thank you for your helpful suggestions and comments. We respond to the comments point by point in order below.

Weakness 1 and Question 1:

1. Error detection by uFusedu_\mathrm{Fused} and Confidence

To evaluate the reliability of uFusedu_\mathrm{Fused}, we compare it with Confidence for detecting incorrect diagnoses on DDR using Qwen3-VL 8B. Both scores are evaluated against the same final diagnostic predictions and ground-truth labels. We sort samples independently by increasing uFusedu_\mathrm{Fused} and decreasing Confidence, then divide each ranking into five approximately equal-sized, non-overlapping quintiles. The first quintile contains the predictions ranked as most reliable by each score. We report the diagnostic error rate within each quintile. In the first quintile, the diagnostic error rate is 2.47% when ranked by uFusedu_\mathrm{Fused} and 17.28% when ranked by Confidence, a difference of 14.81 percentage points.

QuintileError rate by uFusedu_\mathrm{Fused} (%)Error rate by Confidence (%)
1st2.4717.28
2nd4.9419.75
3rd12.3516.05
4th24.6920.99
5th46.2516.25

2. The uncertainty in ablation

To assess how sample dependent evidence construction and conflict based downweighting affect error detection, we compare the full framework with two ablated variants. The first fixes the evidence strength of both pathways to Δ = 1. The second removes concept evidence downweighting (w/o CD). Linear fusion does not define the Dirichlet uncertainty mass used here, so it is excluded from this comparison. For each variant, we sort its final predictions by increasing uFusedu_\mathrm{Fused}, divide them into five approximately equal-sized quintiles, and report the diagnostic error rate within each quintile. The first quintile contains the predictions ranked as most reliable by that variant.

DatasetDerm7ptCORDADDR
QuintileFullΔ = 1w/o CDFullΔ = 1w/o CDFullΔ = 1w/o CD
1st0.001.561.565.068.866.332.476.174.94
2nd1.566.253.1315.1917.7220.254.949.886.17
3rd6.2517.199.3820.5132.0517.958.6417.2813.58
4th25.0031.2528.1335.9034.6235.9022.2228.4020.99
5th40.6325.0039.0638.4619.2334.6247.5035.0045.00

All values are within quintile diagnostic error rates (%). The full framework has a lower error rate than both ablated variants in the first two quintiles on all three datasets. These results show that retaining both designs yields fewer errors among the diagnoses ranked as most reliable at the same coverage.

3. Benefits of the proposed designs

Our framework addresses two challenges in medical image diagnosis with frozen LVLMs: hallucination and bias toward common classes under long-tailed class distributions.

To mitigate hallucination, we design two complementary pathways because incorrect clinical findings predicted by a frozen LVLM can mislead the diagnosis. The concept matching pathway infers a diagnosis from predicted clinical findings and class-conditional concept statistics, while the visual perception pathway predicts the diagnosis directly from the image. Without training a separate evidence estimator, we use debiased class probabilities to allocate evidence across classes and sample-dependent logit gaps to determine the initial strength of that evidence. Because incorrect concept predictions can also have large logit gaps, we downweight concept evidence before fusion when the predicted findings conflict with co-occurrence patterns estimated from annotated data. The fused uncertainty reflects both pathways’ evidence strengths and the disagreement between their diagnostic predictions.

To mitigate long-tail bias, we introduce a separate prior correction strength for each pathway because class bias in a frozen LVLM’s outputs may differ from the bias implied by the empirical class frequencies of the target dataset. The concept-matching pathway adjusts the explicit class prior in its Bayesian diagnosis, while the visual-perception pathway adjusts the class probabilities derived from the frozen LVLM’s logits. This lets us control the degree of prior correction separately for each pathway.

Our experiments and ablation studies support the effectiveness of these designs in improving diagnostic performance and the reliability of uncertainty estimates.

AuthorsOpenReview ↗

Response to Reviewer hEnh [2/3]

[2/3] Due to the character limit, this response is split into three parts. This is part 2/3.

Weakness 2 and Question 3:

1. Number of retrieved examples

We evaluated Qwen3-VL-8B on Derm7pt with the number of retrieved examples per query, TT, ranging from 0 to 8. We used the same training reference pool and test split throughout. At T=0T = 0, we omitted the reference examples from both pathways’ prompts. We report classification performance, inference time per test image, and mean fused uncertainty.

The following table shows that the number of retrieved examples TT from 0 to 8. At T=0T = 0, we removed all reference images and their accompanying demonstration text. BACC increased from 77.92% at T=0T = 0 to 88.09% at T=4T = 4. With more than four examples, both BACC and macro F1 were below their values at T=4T = 4, while inference time per test sample increased from 10.68 s at T=4T = 4 to 48.22 s at T=8T = 8.

TTBACC (%)Macro-F1 (%)Time (s)Uncertainty
077.9278.732.130.2838
179.4078.544.460.2195
282.5681.446.440.1853
385.4684.249.720.1498
488.0985.1310.680.1284
587.3484.8714.980.0973
686.2684.4521.550.1251
784.2481.4429.040.0789
883.6783.2248.220.0727

2. Retrieval quality

To assess how retrieval quality affects diagnosis, we keep the number of reference examples fixed at four on Derm7pt using Qwen3-VL-8B. For each test image, we use the original top-four retrieved examples as the control. We then randomly replace one to four of these examples with images sampled from the same training set outside the original top four. Each replacement image retains its own diagnostic label and concept annotations, and both pathways receive the same set of examples. We repeat the sampling and evaluation with three random seeds, using the same control for all runs. For each nonzero replacement level, we report the mean and sample standard deviation across the three runs for BACC, Macro-F1, and the test-set mean of uFusedu_\mathrm{Fused}. We report a single value for each metric in the shared 0/4 control. Replacing a single retrieved example lowers mean BACC and Macro-F1 relative to the control. Both metrics decline further as more examples are replaced, while mean uFusedu_\mathrm{Fused} generally increases. These results show that the final diagnosis is sensitive to retrieval perturbations, with larger performance drops when more of the original retrieved examples are replaced.

References replacedBACC (%)Macro-F1 (%)Uncertainty
0/488.0985.130.1284
1/485.87 ± 0.9383.34 ± 0.880.1290 ± 0.0027
2/483.01 ± 0.8583.16 ± 0.820.1532 ± 0.0019
3/480.94 ± 0.5778.92 ± 0.750.1508 ± 0.0021
4/477.63 ± 1.0676.85 ± 1.260.2092 ± 0.0017
AuthorsOpenReview ↗

Response to Reviewer hEnh [3/3]

[3/3] Due to the character limit, this response is split into three parts. This is part 3/3.

Question 1:

See the response of weakness 1.

Question 2:

We will clarify which components were replaced or retained in each baseline. Where the visual encoder could be replaced without altering the method’s core mechanisms, we used MedImageInsight. We retained the original backbones when they were integral to the method, as detailed below. For baselines with an LLM component, we replaced the original LLM with Qwen3-VL, CheXagent, or Idefics3 in the corresponding experiments. Apart from these substitutions, we retained each baseline’s original procedure and algorithmic mechanisms.

For BiomedCoOp, we preserved the BiomedCLIP image and text encoders and the original BiomedCLIP-based prompt-learning procedure, including its semantic-consistency constraint, knowledge-distillation strategy, and statistics-based prompt selection. BiomedCLIP is the vision-language backbone on which BiomedCoOp’s prompt-learning method is built, so replacing it would alter a defining component of the method. In each experiment, we replaced the GPT component used to generate text prompts with the corresponding LVLM.

SOTA is a multi-model ensemble method. In its best-performing configuration, we replaced CLIP with MedImageInsight and retained DINOv2, DINOv3, and adaptive optimal transport fusion.

Question 3:

See the response of weakness 2.

Question 4:

Thank you for pointing this out. “Qwen3-VL 7B” is a typo. The checkpoint used in our experiments was Qwen3-VL-8B-Instruct. We will correct the model name to “Qwen3-VL 8B”.

AuthorsOpenReview ↗

Summary of Responses

Thank you for your detailed comments and suggestions. As the discussion period draws to a close, we would like to briefly summarize the additions and revisions addressing your concerns and invite your further feedback.

Regarding Weakness 1 and Question 1 about whether the proposed uncertainty improves error detection over standard confidence measures, we have supplemented our previous confidence comparison and uncertainty ablations with an AUROC comparison between the visual pathway’s own logit-gap confidence and fused uncertainty. For each dataset, we used the backbone achieving the highest BACC in Table 1 and its corresponding parameters. The complete results are as follows:

DatasetAUROC (Visual Confidence)AUROC (uFusedu_{\mathrm{Fused}})Δ AUROC
Derm7pt0.59150.8317+0.2402
CORDA0.65400.7048+0.0508
DDR0.67740.8038+0.1264

Fused uncertainty achieved a higher error-detection AUROC on all three datasets. In addition, the uncertainty ablations provided in our previous response show that the full framework has lower error rates in the two most reliable quintiles than either fixing the evidence strength or removing concept-conflict downweighting across all three datasets.

Regarding Weakness 2 and Question 3, we have added experiments varying the number of retrieved reference examples from 0 to 8, together with retrieval perturbation experiments that keep four reference examples and randomly replace one to four of them. We report classification performance and mean fused uncertainty to examine the effects of reference quantity and quality.

Regarding Question 2, the revised manuscript details the components retained or replaced in each baseline. In particular, for BiomedCoOp, we preserved the BiomedCLIP image and text encoders and the original prompt-learning mechanisms, replacing only the language-model component used to generate text prompts.

Regarding Question 4, we have corrected the model name to Qwen3-VL-8B-Instruct and fixed the previous “7B” typo.

If your time permits, we would greatly appreciate your feedback on whether these additions and revisions address your main concerns and whether any issues require further clarification. Thank you for your time and consideration.

Reviewer_6sh7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a training-free framework for trustworthy medical image diagnosis, integrating concept-based reasoning, visual perception, long-tail debiasing, and Dirichlet evidence fusion to improve diagnostic accuracy and uncertainty estimation. Numerous experiments demonstate the effectiveness of this method.

Critical strengths
  • Clear Presentation: The paper is well-written and easy to follow, allowing readers to quickly grasp the core idea and overall methodology.

  • Interesting Topic: The paper addresses an interesting and practically important problem: improving the trustworthiness of medical image diagnosis while maintaining interpretability in a training-free setting.

Critical weaknesses
  • Reliability of Uncertainty Estimation: The proposed uncertainty estimation relies on the logit gap. However, given that VLMs can produce overconfident incorrect predictions, it is unclear why the logit gap can serve as a reliable uncertainty indicator.

  • Terminology Inconsistency: The paper repeatedly refers to the underlying models as LLMs, although these models accept images as input and are more accurately categorized as VLMs. The authors should clarify this distinction and use consistent terminology throughout the paper.

  • Limited Model Diversity: The experiments are conducted on a limited set of VLMs. I suggest evaluating more recent VLMs of different sizes, including Mixture-of-Experts (MoE) architectures, as all currently evaluated models appear to use dense architectures.

  • Error Propagation Across Stages: The proposed framework relies heavily on the quality of the first-stage retrieval. If irrelevant or misleading examples are retrieved, they may negatively affect both diagnostic pathways in Stage 2. It would be valuable to investigate the robustness of the framework to imperfect retrieval results.

Questions for authors

Please see in the weakness section

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6sh7 [1/2]

[1/2] Due to the character limit, this response is split into two parts. This is part 1/2.

Thank you for your helpful suggestions and comments. We respond to the comments point by point in order below.

Weakness 1:

The frozen LVLMs do indeed tend to produce overly confident logits, but Δ is merely used as a measure of relative logit separation for constructing pathway-specific evidence strength. In our method, what is actually used as an uncertainty indicator is uFusedu_\mathrm{Fused} in Equation 11. Equation 8 and 9 discount concept pathway evidence when the predicted concepts are inconsistent with concept co-occurrence relationships estimated from the reference data. Equation 11 combines both pathways’ evidence with their cross pathway conflict. For fixed evidence strengths, greater disagreement between their diagnostic distributions leads to higher fused uncertainty. Thus, the final score depends on both evidence strength and consistency, rather than the logit gap alone.

To evaluate the reliability of uFusedu_\mathrm{Fused}, we compare it with Confidence for detecting incorrect diagnoses on DDR using Qwen3-VL 8B. Both scores are evaluated against the same final diagnostic predictions and ground-truth labels. We sort samples independently by increasing uFusedu_\mathrm{Fused} and decreasing Confidence, then divide each ranking into five approximately equal-sized, non-overlapping quintiles. The first quintile contains the predictions ranked as most reliable by each score. We report the diagnostic error rate within each quintile. In the first quintile, the diagnostic error rate is 2.47% when ranked by uFusedu_\mathrm{Fused} and 17.28% when ranked by Confidence.

QuintileError rate by uFusedu_\mathrm{Fused} (%)Error rate by Confidence (%)
1st2.4717.28
2nd4.9419.75
3rd12.3516.05
4th24.6920.99
5th46.2516.25

Weakness 2:

Thanks for pointing out the terminology inconsistency. Our diagnostic backbones are built on large language models and take both an image and a text prompt as input, so we will refer to them consistently as large vision-language models (LVLMs) throughout the manuscript.

AuthorsOpenReview ↗

Response to Reviewer 6sh7 [2/2]

[2/2] Due to the character limit, this response is split into two parts. This is part 2/2.

Weakness 3:

To broaden the range of backbone sizes and architectures in our evaluation, we extended our evaluation to Gemma 4 26B-A4B-IT, an instruction-tuned vision-language model released in April 2026 with a mixture-of-experts architecture. It has 25.2B total parameters and activates approximately 3.8B parameters per token. We evaluated our method and CBVLM, KPL, Two-Step with this backbone on Derm7pt, CORDA, and DDR, and report BACC and Macro-F1 for each dataset in the table below. This experiment extends our evaluation beyond the original dense backbones and provides a comparison with CBVLM using the same MoE backbone.

Derm7ptCORDADDR
BACC (%)Macro-F1 (%)BACC (%)Macro-F1 (%)BACC (%)Macro-F1 (%)
Ours84.9780.3674.6371.7962.9258.57
CBVLM80.2478.6273.4469.9860.1255.47
KPL70.7665.4268.1566.8352.6342.71
Two-Step75.6171.9368.6667.3138.6536.48

Weakness 4:

To assess how retrieval quality affects diagnosis, we keep the number of reference examples fixed at four on Derm7pt using Qwen3-VL-8B. For each test image, we use the original top-four retrieved examples as the control. We then randomly replace one to four of these examples with images sampled from the same training set outside the original top four. Each replacement image retains its own diagnostic label and concept annotations, and both pathways receive the same set of examples. We repeat the sampling and evaluation with three random seeds, using the same control for all runs. For each nonzero replacement level, we report the mean and sample standard deviation across the three runs for BACC, Macro-F1, and the test-set mean of uFusedu_\mathrm{Fused}. We report a single value for each metric in the shared 0/4 control. Replacing a single retrieved example lowers mean BACC and Macro-F1 relative to the control. Both metrics decline further as more examples are replaced, while mean uFusedu_\mathrm{Fused} generally increases. These results show that the final diagnosis is sensitive to retrieval perturbations, with larger performance drops when more of the original retrieved examples are replaced.

References replacedBACC (%)Macro-F1 (%)Uncertainty
0/488.0985.130.1284
1/485.87 ± 0.9383.34 ± 0.880.1290 ± 0.0027
2/483.01 ± 0.8583.16 ± 0.820.1532 ± 0.0019
3/480.94 ± 0.5778.92 ± 0.750.1508 ± 0.0021
4/477.63 ± 1.0676.85 ± 1.260.2092 ± 0.0017
AuthorsOpenReview ↗

Summary of Responses

Thank you for your helpful comments and suggestions. As the discussion period draws to a close, we would like to briefly follow up on the additional experiments and revisions addressing your concerns.

Regarding Weakness 1 about the reliability of the logit gap, we further compared the visual pathway’s own logit-gap confidence with the final fused uncertainty for error detection across all three datasets. For each dataset, we used the backbone achieving the highest BACC in Table 1 and its corresponding parameters. Visual confidence is defined as the difference between the largest logit and the mean of the remaining logits, while fused uncertainty is measured by uFusedu_{\mathrm{Fused}}. A higher AUROC indicates better discrimination between correct and incorrect predictions. The complete results are as follows:

DatasetAUROC (Visual Confidence)AUROC (uFusedu_{\mathrm{Fused}})Δ AUROC
Derm7pt0.59150.8317+0.2402
CORDA0.65400.7048+0.0508
DDR0.67740.8038+0.1264

Fused uncertainty achieved a higher error-detection AUROC on all three datasets. As explained in our previous response, the logit gap determines the initial evidence strength of each pathway, while the final uncertainty also incorporates concept-conflict downweighting and cross-pathway evidence fusion.

Regarding Weakness 2, we have consistently adopted LVLMs (large vision language models) throughout the revised manuscript to accurately describe the diagnostic backbones that accept both image and text inputs.

Regarding Weakness 3, we have added comparative experiments using Gemma 4 26B-A4B-IT as a shared MoE backbone, extending the evaluation beyond the original dense architectures.

Regarding Weakness 4, we have added retrieval perturbation experiments: keeping four reference examples, we randomly replace one to four of them and report classification performance and mean fused uncertainty across three random seeds to assess the impact of retrieval quality on the framework.

If your time permits, we would greatly appreciate your feedback on whether these additions and revisions address your main concerns and whether any issues require further clarification. Thanks for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Response to Reviewer qbTU

Thank you for reviewing our response and offering further suggestions.

Uncertainty comparison

Following your suggestion, we compared the visual pathway’s own confidence with our fused uncertainty across all three datasets. We used the backbone that achieved the highest BACC among the three models in Table 1, along with its corresponding parameters. We computed the AUROC for error detection using the visual pathway’s own confidence (the gap between the largest logit and the remaining logits) and the fused uncertainty (uFusedu_\mathrm{Fused}). For error detection, AUROC is the probability that a randomly selected incorrect prediction is ranked as less reliable than a randomly selected correct prediction.

DatasetAUROC (Visual Confidence)AUROC (uFusedu_\mathrm{Fused})Δ\Delta AUROC
Derm7pt0.59150.8317+0.2402
CORDA0.65400.7048+0.0508
DDR0.67740.8038+0.1264

uFusedu_\mathrm{Fused} achieved a higher AUROC than the visual pathway’s own confidence on all three datasets. Derm7pt showed the largest AUROC gain among the three datasets. The pathway ablation results in Table 3 also show that the BACC gain of the full framework over the visual pathway alone was greater on Derm7pt than on either CORDA or DDR.

Per-pathway prior correction

Standard post-hoc methods typically apply class-prior correction to the output of a single classifier. In our design, we apply the correction separately to the two pathways before evidence fusion, accounting for how each pathway produces its diagnostic distribution. The concept pathway explicitly incorporates a class prior in Bayesian inference, so we directly adjust the weight of this prior in diagnostic inference. In the visual pathway, the frozen model directly outputs a score for each diagnostic class, which we adjust using class frequencies estimated from the training set. Setting the correction strength separately for each pathway allows us to adjust the two diagnostic distributions independently before evidence fusion.

Regarding whether the same correction can be applied to the baseline, the first stage of the CBVLM outputs discrete concept labels and does not construct a diagnostic class distribution to which our concept pathway correction could be applied. In the second stage, CBVLM feeds the concept labels and image into the model and extracts a single diagnostic label from the generated text. Our visual pathway correction requires a score for each candidate diagnostic class. Therefore, neither correction can be applied directly to the output of the corresponding stage in the original CBVLM pipeline.

On long tailed tasks such as DDR, we are particularly interested in how well the model identifies disease grades with fewer samples. The new per-class results in Table 8 of Appendix K show that the two methods achieve comparable overall BACC with the same Idefics3 backbone and test set. Our method achieves higher recall and F1 scores than CBVLM on the Mild, Severe, and Proliferative grades. Thus, the advantage we observe on DDR is concentrated in the recognition of these underrepresented grades.

As shown in Table 3, applying long-tail correction to the concept pathway alone raises BACC on DDR to 61.05%. Correcting both pathways further increases BACC to 64.36%, 3.31 percentage points above the best single-pathway correction. This shows that correcting the concept pathway alone accounts for most of the observed improvement and that also correcting the visual pathway further improves BACC within the same evidence fusion framework. Our evidence fusion amplifies the effect of prior correction, and the two components complement each other.

Thank you for confirming that the additional AUROC comparison addresses your concern about uncertainty and for expressing a more favorable overall assessment, leaning toward a borderline score between 2 and 3 after considering the additional evidence.

We greatly appreciate your reconsideration of our paper and understand your remaining reservations about novelty. The more substantial novelty of our work lies in the dual-pathway diagnostic architecture combining concept matching and visual perception, which uses Dirichlet evidence fusion to produce diagnostic predictions and sample-wise uncertainty for hallucination detection and trustworthy diagnosis.

Thanks again for your careful review and constructive suggestions, which have helped us improve the paper and its experimental validation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.