List of Papers Browse by Subject Areas Author List
Abstract
Parameter-efficient adaptation of vision-language foundation models is crucial for precise multimodal understanding of biomedical images, yet existing methods remain deterministic and often struggle under domain shift or ambiguous image-text alignment. This limitation is particularly critical in the clinic, where models should remain robust in low-data regimes and domain shifts. We present Evi-Steer, an evidential cross-modal low-dimensional steering framework for BiomedCLIP that enables uncertainty-aware parameter-efficient fine-tuning while updating only 0.11% of total model parameters. Our approach performs lightweight low-dimensional token updates in both vision and text encoders while simultaneously estimating epistemic uncertainty. These uncertainty estimates update gate residuals, allowing the model to adapt conservatively when evidence is weak. Furthermore, we introduce cross-modal confidence fusion based on Dempster-Shafer theory, enabling visual adaptation to be conditioned on textual confidence and suppressing conflicting or uncertain cross-modal updates. We conduct a comprehensive evaluation on 15 biomedical imaging datasets spanning 8 organs and 8 imaging modalities under few-shot learning and domain generalization settings. Evi-Steer consistently outperforms state-of-the-art methods under few-shot learning and domain shift settings, demonstrating a practical and robust pathway for deploying vision-language models in real-world clinical settings. Code is available at https://github.com/HealthX-Lab/Evi-Steer.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2764_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/HealthX-Lab/Evi-Steer
Link to the Dataset(s)
https://huggingface.co/datasets/TahaKoleilat/BiomedCoOp
BibTex
@InProceedings{KolTah_EviSteer_MICCAI2026,
author = { Koleilat, Taha AND Rivaz, Hassan AND Xiao, Yiming},
title = { { Evi-Steer: Learning to Steer Biomedical Vision-Language Models through Efficient and Generalizable Evidential Tuning } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16879},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper presents Evi-Steer, a parameter-efficient adaptation method for BiomedCLIP that injects small low-rank “steering” modules into the hidden states of both the vision and text encoders (the first d transformer layers), with three ingredients layered on top. First, each injected module outputs both a mean update and an uncertainty channel in a shared r-dimensional latent space, where the uncertainty is turned into Dirichlet-like belief masses. Second, the per-modality belief gates the size of the residual update applied at each layer. Third, textual and visual beliefs are combined via Dempster-Shafer fusion, so that visual adaptation is damped when textual confidence is low. Training uses cross-entropy plus a KL regulariser that pulls concentration parameters toward a Gamma(1,1) prior. The authors report results on nine biomedical few-shot classification datasets (K=4/8/16) and six OOD targets for domain generalisation, claiming consistent gains over CoOp/CoCoOp/KgCoOp/ProGrad/BiomedCoOp/LP++/CLIP-Adapter/Tip-Adapter-F/GDA/CLIP-LoRA while updating only ~0.11% of parameters (221K).
- 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 is well-chosen. Few-shot adaptation of biomedical VLMs with attention to cross-modal confidence is a real gap; most of the CLIP PEFT literature handles the two modalities independently, or folds textual context learning into prompt tokens without any explicit confidence coupling. Tying visual adaptation strength to textual confidence via a fusion rule is a sensible design move, and I can see why it would help on OOD where the text side (essentially fixed clinical class names) should be more reliable than the image side.
-The activation-space “steering” formulation is clean. Eqs. 1-4 are simple enough to follow and the zero-init of W ensures the model degrades gracefully to the pretrained backbone at start-up. The parameter budget is small (0.11% / 221K) and is in the same ballpark as CLIP-LoRA, so the efficiency claim is fair.
-
The experiments are broad including nine datasets for few-shot settings and two domain-shift settings on three target datasets. This is a reasonable effort. Gains on the harder in-distribution datasets (OCTMNIST at K=16: 80.37 vs 75.43 for GDA; RETINA at K=16: 68.53 vs 62.85 for Tip-Adapter-F) are non-trivial.
-
The ablation in Table 3 is structured sensibly by removing each of the four components one at a time. The pattern (visual adaptation matters most, evidential/belief modules matter more for OOD than ID) is consistent with what you would expect from the narrative.
-
Figures 2a/2b give the depth and dimension sweeps, and the fact that r=4 is the optimum rather than r=16 is a useful finding for practitioners.
- 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.
-
A key precedent is missing from the references. The paper’s evidential formulation of softplus, Dirichlet concentration β = e+1, uncertainty u = β^(-1), is the classification-style evidential construction due to Sensoy, Kaplan & Kandemir (NeurIPS 2018, “Evidential Deep Learning to Quantify Classification Uncertainty”). That paper is not cited. Instead ref [2] is Amini et al. 2020 on Deep Evidential Regression, which uses a Normal-Inverse-Gamma prior and is a different construction. Ref [8] is a 2025 survey. The direct precedent is Sensoy et al. , and the paper should cite it both to credit prior work and because it would let the reader check whether what Evi-Steer does matches the standard Dirichlet/subjective-logic framework. (It doesn’t quite: Sensoy’s uncertainty is u = K/S at the class level, where K is the number of classes and S is the sum of concentrations. Evi-Steer applies u = β^(-1) element-wise to a latent tensor, not over classes. This differs from the standard evidential interpretation and should be justified why is this different. )
-
The Dempster–Shafer framing is heavier than the maths warrants. Eq. 9 is a two-hypothesis DST combination with a frame {H1, H0} = {apply update, ignorance}, giving m(H1) = m_t^{H1} · m_v^{H1} / (1 − conflict). This reduces to the product of two scalar confidences divided by a near-constant normaliser; there is no conflict mass tracked beyond the denominator, no support for composite hypotheses, and nothing distinguishing this from, say, a sigmoid-gated product of two confidence scores. The paper is free to call this DST fusion (technically it is), but the motivation in the text (“Unlike standard probabilistic fusion, DST separates confidence from uncertainty. . . “) oversells the mathematical content. An ablation replacing Eq. 9 with a plain product gate m_t * m_v would actually settle whether the DST language is doing any work, and I would encourage the authors to run that comparison.
-
Importantly, a directly relevant baseline is omitted. Ref [19] (Koleilat et al. , CLIP-SVD) is cited in the related work but not in any of the results tables. CLIP-SVD is in the same problem space (parameter-efficient adaptation of CLIP/BiomedCLIP, few-shot biomedical benchmarks, ~0.04% of parameters) and its reported numbers on overlapping datasets are strong. Without this comparison, the claim of “consistently outperforms state-of-the-art methods” cannot be verified.
-
The “15 datasets” claim is overbroad. Few-shot learning uses 9 datasets; domain generalisation uses 2 source + 6 ‘target’ datasets, but for domain generalisation the target datasets can only ever be evaluated zero-shot, not trained on. So, the framing “comprehensive evaluation on 15 biomedical imaging datasets . . “ in the abstract conflates these two settings. I would state it more cleanly: nine few-shot datasets, two source-target DG setups instead of 15.
-
BiomedCoOp numbers don’t match the BiomedCoOp paper. At K=16, this paper reports BiomedCoOp average accuracy as 77.15; the BiomedCoOp CVPR 2025 paper reports 72.42 ± 1.69 on 11 datasets. At K=8 the gap is 72.43 here vs 68.32 there. The obvious explanation is the dataset. In particular, this paper uses 9 datasets, BiomedCoOp used 11 (DermaMNIST and KneeXray are the likely excluded ones, both historically harder). This is fine as long as it’s stated: please explicitly report which datasets were dropped and why, and ideally rerun on the full BiomedCoOp benchmark set so the comparison is apples-to-apples.
-
No standard deviations in Tables 1 and 2.The paper says results are averaged over three runs, but no error bars are reported unlike BiomedCoOp’s Table 1, which reports mean ± std. Several of the gaps over the best baseline are in the 1-3 percentage point range (e. g. K=16 BUSI: 73.16 vs 71.89; K=16 COVID: 73.44 vs 78.72 where Evi-Steer actually loses by 5 points; BTMRI K=16: 86.39 vs 83.30). With three seeds, fold-to-fold variance is almost certainly in this range. Please add standard deviations and, ideally, a paired test on the per-dataset accuracies. This will allow the reviewers to see the finer details of the results to verify the claims.
-
Evi-Steer does not uniformly outperform at K=16.Table 1 shows Evi-Steer trailing GDA on COVID (73.44 vs 75.65 - actually the listed COVID value for 16-shot is lower than 8-shot which is not intuitive in my experience), trailing BiomedCoOp on COVID (78.72), trailing TIP-Adapter-F on LC25000 (92.35 vs 92.05). The “consistently achieves the best overall performance” claim is therefore “best on average” rather than “best everywhere,” and should be phrased that way. Also, Evi-Steer’s COVID number drops from 74.69 at K=8 to 73.44 at K=16, which is a non-monotonic result worth commenting on.
-
ReFT precedent should be discussed. The paper cites Wu et al. (ReFT, NeurIPS 2024) as ref [36] and correctly attributes the “representation steering paradigm. “ But ReFT’s LoReFT is exactly a low-rank subspace idea on frozen hidden states via down/up projections, which is structurally the same as Eqs. 1-4 here. The novelty of Evi-Steer relative to ReFT is the uncertainty gating and the cross-modal DST fusion, not the “low-dimensional activation-space update” which is LoReFT’s contribution. The paper should make that delineation explicit in the contributions paragraph rather than presenting the Latent Dimension Representation itself as novel.
-
Missing details on the uncertainty channel: Eq. 1 projects T^(n) into [Z^μ, Z^σ] ∈ R^{C × P_t × r}, and Eq. 5 maps Z^σ through softplus-squared. I cannot tell from the text whether the same down-projection W(\downarrow)_t is shared between the mean and uncertainty paths (i. e. whether W(\downarrrow)_t outputs 2r channels that are split), or whether there are separate projections for each. Please clarify.
-
λ_KL = 10^-4 is fixed. No sensitivity analysis on this regulariser weight is shown. Given that the claim is “evidential regularisation produces statistically meaningful evidence,” this really wants an ablation of λ_KL ∈ {0, 10^-5, 10^-4, 10^-3, 10^-2} to show the regulariser is actually doing what’s claimed. Without that, “w/o Evidential Update” in Table 3 only shows removing the entire uncertainty-gating mechanism matters; it doesn’t isolate the KL regulariser.
Minor citation issue. Ref [38] lists CLIP-LoRA as “Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition”. It’s actually a CVPR Workshops paper (CVPRW 2024). Worth fixing.
- Minor numerical inconsistency. Abstract reports average accuracies of 71.45% / 77.34% / 81.21% at K=4/8/16; Table 1 gives 71.43 / 77.33 / 81.18 (I recomputed and the table values are the correct means of the per-dataset rows). The abstract should match the table.
-
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The method is reasonable but incremental over ReFT (which provides the steering template) plus Sensoy-style evidential classification (which provides the uncertainty construction). The empirical case is weakened by a missing direct baseline (CLIP-SVD), a dataset discrepancy that isn’t explained, absence of error bars across three-seed runs which makes it difficult to verify the actual improvements of the results, and non-uniform performance of the method where the narrative suggests that it outperforms uniformly at K=16.The DST framing, while not incorrect, adds less than the method section implies. Together these hold the paper below the bar for me at this venue.
- Reviewer confidence
Very confident (4)
- [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 introduces Evi-Steer, an evidential cross-modal low-dimensional steering framework for parameter-efficient adaptation of BiomedCLIP. The core idea is to inject lightweight adapters into both vision and text encoders that operate in the activation space (rather than the weight space), estimate epistemic uncertainty via evidential modeling, and use Dempster–Shafer belief fusion to condition visual adaptation on textual confidence. The method updates only 0.11% of model parameters and is evaluated on 15 biomedical datasets under few-shot and domain generalization settings.
- 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 combination of activation-space representation steering with evidential uncertainty gating and DST cross-modal fusion is novel. No prior work integrates all three for VLM adaptation, particularly in the biomedical domain.
2.Evaluation breadth is a clear strength. 15 datasets spanning 8 organs and 8 modalities, tested under both few-shot (K=4/8/16) and domain generalization protocols, with 10 baseline comparisons.
3.The ablation study (Table 3) is well-designed and informative. Removing each component shows measurable degradation, and the observation that evidential/DST modules have a larger OOD impact than ID impact is a meaningful insight, it directly supports the clinical motivation for uncertainty-aware adaptation.
4.The parameter efficiency claim is well-supported. At 221K trainable parameters (0.11%), the method matches CLIP-LoRA’s budget while consistently outperforming it, which strengthens the practical deployment argument.
5.Domain generalization results (Table 2) are convincing. Evi-Steer achieves the best OOD transfer on both breast ultrasound and brain MRI domains, which is the most clinically relevant evaluation setting.
- 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 individual technical components are not new. Activation-space steering follows directly from ReFT (Wu et al., NeurIPS 2024), evidential uncertainty via Dirichlet concentrations is standard (Amini et al., NeurIPS 2020), and DST fusion is classical. The novelty lies in the combination, which is valid but should be acknowledged more explicitly.
2.Several closely related works are missing from the discussion. CILMP (Du et al., IEEE TMI 2025) applies ReFT-style low-rank linear subspace interventions to generate disease-specific prompts for medical VLM classification, the same core mechanism. EviVLM (Pan et al., IEEE TMI 2025) integrates evidential learning with VLMs for medical image segmentation using DST combination. Bayesian-PEFT (NeurIPS 2024) combines parameter-efficient fine-tuning with evidential deep learning for vision foundation models. These are critical references that the paper must cite and differentiate from.
3.The marginal gains from the evidential and DST modules are modest in the ablation. Removing the evidential update costs only 1.10% HM and removing cross-modal belief costs 0.68% HM. While the OOD degradation is larger, the overall improvement may not fully justify the added complexity of evidential modeling and DST fusion over a simpler confidence-free adapter.
4.The citation quality needs significant refinement. Three references ([24], [27], [30]) share the same title, “Brain tumor mri dataset”, and are differentiated only by a Kaggle uploader’s name and year. Reference [27] cites a single first name (“Pradeep”) with no surname, no URL, and no DOI. Reference [30] (“Sherif, M.M.”) follows the same pattern.
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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
The overall quality of the bibliography falls below MICCAI standards.
- 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 presents a well-motivated and technically sound approach that addresses a real gap: uncertainty-aware parameter-efficient adaptation of biomedical VLMs. The evaluation is comprehensive and the results are consistently strong, particularly under domain shift. The formulation is clean and the core methodology is well-explained. However, the novelty is primarily in the combination of existing techniques (ReFT + evidential learning + DST), and the marginal ablation gains from the uncertainty modules raise questions about whether the added complexity is justified. The missing discussion of closely related works (CILMP, EviVLM, Bayesian-PEFT) is a notable gap.
- 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
This paper proposes Evi-Steer, a parameter-efficient adaptation framework for biomedical vision–language models (VLMs), instantiated on BiomedCLIP. The method inserts low-dimensional, cross-modal representation steering modules into early transformer layers and computes token-wise, evidential uncertainty in the latent space to gate residual updates, fusing text–vision confidence via Dempster–Shafer theory. Across 15 datasets, Evi-Steer reports consistent gains over prompt- and adapter-based baselines in few-shot learning and under domain generalization, while training only 0.11% of parameters.
- 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 a clinically important problem by improving parameter-efficient adaptation of biomedical vision-language models. 2.The proposed Evi-Steer framework is technically novel in combining low-dimensional representation steering with evidential uncertainty estimation. 3.The uncertainty-aware gating mechanism is well motivated, allowing the model to adapt conservatively when evidence is weak and reducing unreliable feature updates. 4.The experimental evaluation is comprehensive, covering 15 biomedical datasets across multiple organs, modalities, few-shot regimes, and domain generalization scenarios.
- 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.Although Dempster–Shafer fusion is introduced as a key component, the paper provides limited analysis of why this fusion strategy is preferable to simpler alternatives. 2.Lack ablation study of parameter λ KL
- 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.
- 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?
The paper is clear clinical relevance, technical novelty, and strong empirical results across diverse biomedical datasets. The proposed uncertainty-aware adaptation framework is promising and effective. Although there are some incomplete aspects, given the novelty of the method and the breadth of the experiments, I suggest accept.
- Reviewer confidence
Very confident (4)
- [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 received two positive reviews (1 A and 1 WA) and one WR. Reviewers generally agree on the appropriateness of the problem, the novelty of the method, at least in the combination of different modules, the extensive empirical validation, the strong results, the clinical relevance, and the adequacy of the ablation studies.
Regarding Reviewer 1’s main reasons for recommending weak reject, the missing quantitative comparison with CLIP-SVD is not considered critical, as it was an arXiv preprint at the time of the submission deadline. The perceived incrementality over ReFT and concerns about the DST narrative appear somewhat subjective. Other experimental inconsistencies raised by Reviewer 1 seem mostly solvable with minor modifications.
Two reviewers also noted the missing sensitivity analysis of λ_KL. From a correctness perspective, this is acceptable as long as the value is fixed across datasets and few-shot regimes, which appears to be the case. Nonetheless, such an analysis would have been a good addition. However, space constraints should also be considered.
Overall, the positive aspects outweigh the concerns, which are largely addressable with minor revisions. I urge the authors to carefully address the reviewers’ comments, correct minor issues, and include a discussion of missing references to further strengthen the final manuscript.
