Pith. sign in

REVIEW 3 major objections 5 minor 29 references

This paper proposes MoBE, a training-free test-time adaptation method that routes medical vision-language models to the right modality experts using predictive entropy and refines them with online Bayesian statistics, reporting average accu

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

A training-free test-time adaptation method (MoBE) routes between modality experts by entropy and adapts their prototypes/priors online, improving medical VLM accuracy by 4.3–7.2 points across benchmarks.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Clear optimization-free TTA method for medical VLMs with a solid ablation story, but the headline gains hinge on unreported per-dataset hyperparameters and single-run numbers. the 3 major comments →

arxiv 2607.16726 v1 pith:GQQJWON5 submitted 2026-07-18 cs.CV

Can Experts Adapt Without Training? On Test-Time Modality Generalization in MVLMs

Raza Imam , Darakshan Rashid , Yutong Xie , Dwarikanath Mahapatra , Brejesh Lall , Mohammad Yaqub This is my paper
classification cs.CV
keywords test-time adaptationmixture of expertsmedical vision-language modelsmodality generalizationentropy routingBayesian adaptationBiomedCLIPoptimization-free inference
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that medical vision-language models can generalize to unseen imaging modalities at test time without any gradient updates, by treating multiple modality-specialized experts as a mixture and routing each test sample to the experts whose predictive uncertainties are similar. It identifies a specialization-generalization dilemma: uniformly averaging all experts dilutes modality-specific knowledge, while committing to a single most-confident expert risks severe mismatch under distribution shift. To resolve this, the paper proposes MoBE, which combines entropy-based dynamic routing with per-expert Bayesian adaptation: each selected expert maintains online class prototypes and class priors that are updated only when the expert's prediction is confident. Across seen, unseen, and heterogeneous medical benchmarks, MoBE reports average accuracy gains of +4.72, +7.17, and +4.3 points over previous test-time adaptation methods, with no parametric updates and no backpropagation.

Core claim

MoBE's central claim is that a frozen mixture-of-experts medical vision-language model can route-and-adapt at test time without optimization. Dynamic-k entropy routing selects experts whose entropies fall within a threshold gap of the most confident expert, then weights them inversely by exponential entropy; Expert Bayesian Adaptation (EBA) updates per-expert class prototypes and class priors through confidence-gated exponential moving averages over the test stream. Because routing and adaptation operate on accumulated statistics rather than model weights, the entire pipeline is optimization-free, and it outperforms both gradient-based and cache-based test-time adaptation methods on the eval

What carries the argument

The load-bearing mechanism is the pairing of dynamic-k entropy routing with Expert Bayesian Adaptation (EBA). Routing uses the relative predictive entropy gap (Eq. 1) to select a variable number of experts and inverse-exponential-entropy weights (Eq. 2) to combine them, which is how the method navigates the specialization-generalization dilemma. EBA maintains per-expert class prototypes, initialized from text embeddings, and per-expert class-prior vectors, updated only when posterior confidence exceeds thresholds θ1 and θ2 (Eqs. 4–5); the adapted per-expert posteriors (Eq. 3) combine prototype similarity with the accumulated prior and are blended with routed logits via λ. The two-stage mecha

Load-bearing premise

The method assumes that a confident prediction under an unseen modality is trustworthy enough to be used as ground truth for updating that expert's prototypes and priors; the paper's own failures on OrganAMNIST and BloodMNIST illustrate that this trust can break down when expert coverage or class separability is poor.

What would settle it

Feed a test stream from a modality unseen by all experts, where one expert is confidently but systematically wrong on a class (for example, confusing colon pathology with normal tissue), and measure whether MoBE's accuracy declines over the stream compared to the frozen backbone; if the confidence-gated updates convert those errors into corrupted prototypes, accuracy should drop as the stream length grows.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Test-time adaptation need not touch model weights: routing decisions plus Bayesian statistics suffice to surpass gradient-based TTA methods on seen modalities.
  • On unseen modalities, the gain is largest (+7.17 over the prior MoME method), suggesting uncertainty-guided routing transfers to modalities never seen by the experts during pretraining.
  • MoBE improves accuracy across different frozen backbones, indicating the mechanism is backbone-agnostic.
  • Being optimization-free, MoBE avoids backpropagation overhead; its main cost is extra forward passes over multiple experts, which the paper flags as a target for future speedups.
  • The dynamic-k selection matches the best fixed top-k while requiring no manual tuning of k.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If entropy-based routing is truly close to oracle cross-entropy routing, the same signal could be applied to test-time ensembling in general vision-language models beyond the medical domain, without any training.
  • The confidence-gated updates suggest a testable extension: replacing hard thresholds with soft reliability weights based on inter-expert agreement might prevent a confidently-wrong expert from corrupting online statistics.
  • The paper's own exceptions (OrganAMNIST, BloodMNIST) define a boundary condition: when expert-bank coverage of the target modality is thin, entropy routing loses discriminability; a fallback to the frozen zero-shot head or an adaptive expert-bank expansion could extend coverage.
  • A natural stress-test is label shift: if the test stream's class distribution drifts, the prior-update mechanism may chase a moving target; a small calibration stream could anchor the prior.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes MoBE, an optimization-free test-time adaptation method for medical vision-language models (MVLMs) with modality-specialized experts. MoBE consists of (i) dynamic-k entropy-gated expert routing, (ii) inverse-entropy expert weighting, (iii) expert-wise Bayesian adaptation that updates class prototypes and priors via confidence-gated running averages, and (iv) a final fusion of gated and adapted logits. The authors report average accuracy gains over state-of-the-art TTA methods of +4.72 on seen modalities, +7.17 on unseen modalities, and +4.3 on heterogeneous medical benchmarks, and claim that training-free expert adaptation can robustly handle modality shifts. The paper also includes ablations on routing, dynamic-k selection, weighting, and backbone transferability.

Significance. If the reported gains are reproducible and the benchmark comparison is fair, MoBE would be a useful contribution: it avoids backpropagation, is conceptually simple, and combines dynamic expert routing with online Bayesian statistics in a way that is not covered by existing cache-based or gradient-based TTA methods. The authors also provide ablations that support each component and show consistency across multiple backbones. However, the central claim is empirical and depends on the exact test-time hyperparameter choices and on consistent aggregation; the manuscript currently does not provide enough information to verify either.

major comments (3)
  1. [§3.1] The sentence 'Dataset-specific hyperparameter values for MoBE are available at Github' is the only source of the exact configuration; no URL, no values, and no validation protocol are given. MoBE has at least five tunable quantities per dataset: τ in Eq. (1), the inverse-entropy weighting in Eq. (2), τ_EBA in Eq. (3), θ1/θ2 in Eqs. (4)–(5), and λ in the final fusion. All tables report one run only. If these values were chosen per test set, the comparison to fixed-hyperparameter baselines (TENT, TDA, MoME) is not apples-to-apples, and the headline gains are not established. Please report the exact hyperparameter values for each dataset, describe how they were selected (e.g., validation split, fixed defaults), and provide error bars or multiple-seed results.
  2. [Table 4] The columns Seen-Avg and Unseen-Avg are not defined, and the numbers are inconsistent with Tables 1–2. For BiomedCLIP, Table 4 reports Seen-Avg=35.45 and Unseen-Avg=23.01, whereas the corresponding averages in Tables 1 and 2 are 33.00 and 20.49. More seriously, +MoBE reports Unseen-Avg=57.35, while Table 2 gives MoBE average accuracy as 38.16. This discrepancy directly affects the claim that MoBE transfers across backbones, and the table should be corrected and its aggregation rule stated.
  3. [§2.2, Eqs. (4)–(5)] Confidence-gated adaptation treats the current posterior as a pseudo-label for updating prototypes and priors. Under an unseen-modality shift, a confidently wrong expert can move its prototype and prior away from the true distribution, and because the updates are sequential and the counters c1,e,c, c2,e,c grow slowly at first, the corruption can compound. The manuscript provides no analysis of whether θ1/θ2 prevent this, no diagnostic of agreement between high-confidence predictions and ground truth, and no comparison against a no-adaptation or lower-confidence variant. This is a concrete correctness-risk concern for the adaptation component and should be addressed either analytically or empirically.
minor comments (5)
  1. [Abstract/§3.1] The code/dataset link is given as a bare 'Github' placeholder; an actual URL is needed.
  2. [Fig. 1b] The caption claims a Pearson correlation between entropy routing and oracle cross-entropy routing, but no correlation coefficient or value is reported. Please provide the actual number and the datasets used.
  3. [Tables 1–2] The naming of 'seen' vs 'unseen' modalities should be made precise relative to the ROCOV2 expert partition. For example, DermaMNIST (dermatoscope) and PathMNIST (colon pathology) are called unseen in Table 2, but the relationship of these modalities to the five expert modalities (CT, MRI, X-ray, ultrasound, angiogram) should be stated explicitly.
  4. [§3.1] The method processes each test sample independently with batch size 1, but the Bayesian updates are sequential (Eqs. 4–5). This makes the results order-dependent. Please clarify whether results are averaged over multiple random test-stream orders or whether a single fixed order was used.
  5. [Throughout] The abbreviation 'MOME' appears in Table 2 while 'MoME' is used elsewhere; please standardize the capitalization.

Circularity Check

0 steps flagged

No significant circularity; MoBE's claims are empirical benchmark results, not derived from their inputs.

full rationale

MoBE is an optimization-free test-time adaptation method and the paper's central claims (+4.72/+7.17/+4.3 accuracy gains) are measured against external MedMNIST/Med-VTAB and heterogeneous benchmarks. There is no derivation chain in which an output quantity is defined in terms of the quantity it is claimed to predict. Eq. 1 defines the dynamic expert set from entropy gaps; Eq. 2 weights experts; Eqs. 3-5 define Bayesian prototype/prior updates from confident predictions. Although the confidence-gated updates are a self-training-style feedback loop, they are not circular by construction: the updated prototypes incorporate new test features h and are not constrained to reproduce the previous logits, and the final fusion is an empirical combination. The paper also benchmarks against external methods (TENT, TDA, MoME) and reports ablations over components, fixed-k vs dynamic-k, and weighting schemes, which provide independent grounding. The main review concerns—unreported per-dataset hyperparameter values and single-run results—are evaluation transparency/testability issues rather than instances of a prediction reducing to a fitted input; there is no quoted evidence that hyperparameters were selected on the test sets. Table 4's seen/unseen averages not matching Tables 1-2 is likewise a reporting inconsistency, not circularity. Self-citations [4,5,12] motivate design choices but are not load-bearing in a logical derivation. Therefore no circularity steps are identified; score 0.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim rests on several unreported per-dataset hyperparameters and on trusting the confidence-gated self-training loop. No new physical entities are posited; the prototypes, priors, and counters are algorithm state rather than invented entities.

free parameters (6)
  • τ (dynamic-k entropy gap threshold) = not reported (per-dataset)
    Controls which experts are selected in Eq. 1; if tuned per benchmark, the reported gains are partly fitted.
  • λ (final fusion blend weight) = not reported
    Controls adaptation strength in s_final = (1-λ)s + λ s_EBA; per-dataset.
  • τ_EBA (prototype similarity temperature) = not reported
    Scales u_c = τ_EBA · h^⊤ μ_e,c in the EBA prediction; per-dataset.
  • θ1 (prototype update confidence threshold) = not reported
    Gates Eq. 4 prototype updates; per-dataset.
  • θ2 (prior update confidence threshold) = not reported
    Gates Eq. 5 prior updates; per-dataset.
  • E (number of modality experts) = 5
    Architecture choice aligned with five ROCOV2 modality groups; not tuned on target benchmarks but affects routing behavior and compute.
axioms (4)
  • domain assumption Entropy is a reliable proxy for expert routing quality: lower H(p_e) means better modality alignment.
    Justified only by Fig. 1b correlation on one paired experiment; this premise underlies dynamic-k selection in Eq. 1.
  • domain assumption High-confidence posterior P(y|x) can serve as a pseudo-label for prototype and prior updates.
    Eqs. 4–5 update on argmax when max P exceeds θ1/θ2; if confident predictions are wrong, the online statistics drift.
  • domain assumption Pretrained text embeddings w_c are valid initial class prototypes in the visual feature space.
    Sec 2.2 B initializes μ_e,c ← w_c/||w_c||; assumes text-visual alignment transfers to the expert MLP outputs.
  • domain assumption Modality experts trained on ROCOV2 remain informative for MedMNIST, Med-VTAB, and heterogeneous medical datasets.
    Sec 3.1 derives experts from ROCOV2 modality groups; the assumption is explicitly weak for BloodMNIST, where MoBE underperforms MoME.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Can Experts Adapt Without Training? On Test-Time Modality Generalization in MVLMs." pith.science (2026). https://pith.science/paper/GQQJWON5

@misc{pith2026260716726,
  author       = {Pith},
  title        = {Pith review of: Can Experts Adapt Without Training? On Test-Time Modality Generalization in MVLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GQQJWON5}},
  note         = {Machine review of arXiv:2607.16726}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Medical vision-language models (MVLMs) promise broad zero-shot generalization, yet their reliability collapses when confronted with unseen modalities and domains, precisely where clinical robustness matters most. To address this gap, we revisit test-time modality generalization from the perspective of Mixture-of-Experts (MoE) and ask: can experts route-and-adapt without any optimization during inference? We identify a fundamental specialization-generalization dilemma at test time, where blindly aggregating modality experts dilutes modality-specific knowledge, while selecting one highly confident expert risks mismatch under shift. To address this, we propose MoBE: a fully optimization-free framework that performs dynamic expert selection and adaptation at test time. MoBE combines entropy-guided dynamic routing in MoE settings with expert-wise Bayesian adaptation, enabling experts to update their confidence and adapt online without gradient updates. Without parametric updates, MoBE augments a static MVLM with test-time routing and online statistics, achieving average accuracy gains of +4.72, +7.17, and +4.3 over state-of-the-art TTA methods across seen, unseen, and heterogeneous medical benchmarks, highlighting the effectiveness of training-free expert adaptation for robust modality generalization.

Figures

Figures reproduced from arXiv: 2607.16726 by Brejesh Lall, Darakshan Rashid, Dwarikanath Mahapatra, Mohammad Yaqub, Raza Imam, Yutong Xie.

Figure 1
Figure 1. Figure 1: (a) The specialization-generalization dilemma induced by uniform expert fusion v/s single-expert reliance at test time. (b) Routing by self-entropy closely matches oracle cross-entropy routing, as they positively correlate. (c) Despite correct routing, pretrained class centroids can misalign with test features, motivating online alignment. partition representation learning across modalities. Mixture-of-Exp… view at source ↗
Figure 2
Figure 2. Figure 2: MoBE Test-Time Adaptation. (A) Frozen BiomedCLIP encoder processes an incoming stream of test images. (B) Dynamic-k entropy routing selects the most reliable modality experts from parallel MLP heads. (C) Selected experts are inverse￾entropy weighted and refined by EBA through confidence-gated prototype and prior updates using the test stream. (D) Final predictions blend routed logits with Bayesian￾adapted … view at source ↗
Figure 3
Figure 3. Figure 3: (a) Component-wise performance: Each component of MoBE improves gen￾eralization. (b) Expert selection effect: MoBE’s dynamic-k routing matches the best fixed top-k while remaining fully automatic. (c) Expert weighting effect: MoBE’s inverse-exponential entropy outperforms averaging and softmax as gating functions. Results on Heterogeneous Dataset: Tab. 3 extends evaluation to hetero￾geneous datasets (e.g.,… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

29 extracted references · 6 linked inside Pith

  1. [1]

    NPJ Digital Medicine8(1), 727 (2025) 1

    Cheng, Z., Ong, A.Y., Wagner, S.K., Merle, D.A., Ju, L., Zhang, H., Chen, R., Pang, L., Li, B., He, T., et al.: Understanding the robustness of vision-language models to medical image artefacts. NPJ Digital Medicine8(1), 727 (2025) 1

    2025

  2. [2]

    arXiv preprint arXiv:2409.19375 (2024) 4

    Han, Z., Yang, J., Wang, G., Li, J., Xu, Q., Shou, M.Z., Zhang, C.: Dota: Distributional test-time adaptation of vision-language models. arXiv preprint arXiv:2409.19375 (2024) 4

    arXiv 2024

  3. [3]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Hanif, A., Shamshad, F., Awais, M., Naseer, M., Khan, F.S., Nandakumar, K., Khan, S., Anwer, R.M.: Baple: Backdoor attacks on medical foundational models using prompt learning. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 443–453. Springer (2024) 1

    2024

  4. [4]

    In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)

    Imam, R., Gani, H., Huzaifa, M., Nandakumar, K.: Test-time low rank adaptation via confidence maximization for zero-shot generalization of vision-language mod- els. In: 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV). pp. 5449–5459. IEEE (2025) 2, 4

    2025

  5. [5]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Imam, R., Hanif, A., Zhang, J., Dawoud, K.W., Kementchedjhieva, Y., Yaqub, M.: Noise is an efficient learner for zero-shot vision-language models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 5820–5829 (2025) 2

    2025

  6. [6]

    Imam, R., Marew, R., Yaqub, M.: On the robustness of medical vision-language models: Are they truly generalizable? In: Annual Conference on Medical Image Understanding and Analysis. pp. 233–256. Springer (2025) 1 10 Imam et al

    2025

  7. [7]

    In: International Conference on Medical Image Computing and Computer-Assisted Intervention

    Ju, L., Zhou, S., Zhou, Y., Lu, H., Zhu, Z., Keane, P.A., Ge, Z.: Delving into out- of-distribution detection with medical vision-language models. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 133–143. Springer (2025) 1

    2025

  8. [8]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Karmanov, A., Guan, D., Lu, S., El Saddik, A., Xing, E.: Efficient test-time adap- tation of vision-language models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 14162–14171 (2024) 2

    2024

  9. [9]

    In: Interna- tional Workshop on Uncertainty for Safe Utilization of Machine Learning in Med- ical Imaging

    Khan, U., Nawaz, U., Sheikh, T.T., Hanif, A., Yaqub, M.: Guardian: Guarding against uncertainty and adversarial risks in robot-assisted surgeries. In: Interna- tional Workshop on Uncertainty for Safe Utilization of Machine Learning in Med- ical Imaging. pp. 59–69. Springer (2024) 1

    2024

  10. [10]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Koleilat, T., Asgariandehkordi, H., Rivaz, H., Xiao, Y.: Biomedcoop: Learning to prompt for biomedical vision-language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 14766–14776 (2025) 6, 7

    2025

  11. [11]

    IEEE Access11, 39351–39372 (2023) 1

    Korevaar, S., Tennakoon, R., Bab-Hadiashar, A.: Failure to achieve domain invari- ance with domain generalization algorithms: An analysis in medical imaging. IEEE Access11, 39351–39372 (2023) 1

    2023

  12. [12]

    arXiv preprint arXiv:2403.07366 (2024) 4

    Lee, J., Jung, D., Lee, S., Park, J., Shin, J., Hwang, U., Yoon, S.: Entropy is not enough for test-time adaptation: From the perspective of disentangled factors. arXiv preprint arXiv:2403.07366 (2024) 4

    Pith/arXiv arXiv 2024

  13. [13]

    arXiv preprint arXiv:2206.04046 (2022) 2

    Li, B., Shen, Y., Yang, J., Wang, Y., Ren, J., Che, T., Zhang, J., Liu, Z.: Sparse mixture-of-experts are domain generalizable learners. arXiv preprint arXiv:2206.04046 (2022) 2

    Pith/arXiv arXiv 2022

  14. [14]

    arXiv preprint arXiv:2404.12876 (2024) 6

    Mo, S., Luo, X., Wang, Y., Li, D.: A large-scale medical visual task adaptation benchmark. arXiv preprint arXiv:2404.12876 (2024) 6

    Pith/arXiv arXiv 2024

  15. [15]

    arXiv preprint arXiv:2509.10025 (2025) 2

    Nikolic, S., Oguz, I., Psaltis, D.: Exploring expert specialization through unsuper- vised training in sparse mixture of experts. arXiv preprint arXiv:2509.10025 (2025) 2

    Pith/arXiv arXiv 2025

  16. [16]

    In: International conference on machine learning

    Niu, S., Wu, J., Zhang, Y., Chen, Y., Zheng, S., Zhao, P., Tan, M.: Efficient test- time model adaptation without forgetting. In: International conference on machine learning. pp. 16888–16905. PMLR (2022) 7

    2022

  17. [17]

    Advances in neural information processing systems37, 48015–48044 (2024) 7

    Osowiechi, D., Noori, M., Vargas Hakim, G., Yazdanpanah, M., Bahri, A., Cher- aghalikhani, M., Dastani, S., Beizaee, F., Ayed, I., Desrosiers, C.: Watt: Weight average test time adaptation of clip. Advances in neural information processing systems37, 48015–48044 (2024) 7

    2024

  18. [18]

    Seco de Herrera, A., et al.: Rocov2: Radiol- ogy objects in context version 2, an updated multimodal image dataset

    Rückert, J., Bloch, L., Brüngel, R., Idrissi-Yaghir, A., Schäfer, H., Schmidt, C.S., Koitka, S., Pelka, O., Abacha, A.B., G. Seco de Herrera, A., et al.: Rocov2: Radiol- ogy objects in context version 2, an updated multimodal image dataset. Scientific Data11(1), 688 (2024) 6

    2024

  19. [19]

    Advances in Neural Information Processing Systems35, 14274–14289 (2022) 2, 7

    Shu, M., Nie, W., Huang, D.A., Yu, Z., Goldstein, T., Anandkumar, A., Xiao, C.: Test-time prompt tuning for zero-shot generalization in vision-language models. Advances in Neural Information Processing Systems35, 14274–14289 (2022) 2, 7

    2022

  20. [20]

    arXiv preprint arXiv:2601.03577 (2026) 2

    Su, Y., Liu, Y.: Variational inference, entropy, and orthogonality: A unified theory of mixture-of-experts. arXiv preprint arXiv:2601.03577 (2026) 2

    arXiv 2026

  21. [21]

    arXiv preprint arXiv:2006.10726 (2020) 7

    Wang, D., Shelhamer, E., Liu, S., Olshausen, B., Darrell, T.: Tent: Fully test-time adaptation by entropy minimization. arXiv preprint arXiv:2006.10726 (2020) 7

    Pith/arXiv arXiv 2006

  22. [22]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Wang, H., Yu, Y., Zheng, H., Zhang, T.: Test-time adaptation of medical vision- language models with mixture of modality experts. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 4649–4658 (2025) 1, 2, 3, 4, 6, 7 Test-Time Modality Generalization 11

    2025

  23. [23]

    In: Proceedings of the 16th ACM International Con- ference on Bioinformatics, Computational Biology, and Health Informatics

    Wang, X., Yang, C.: Moe-health: A mixture of experts framework for robust multi- modal healthcare prediction. In: Proceedings of the 16th ACM International Con- ference on Bioinformatics, Computational Biology, and Health Informatics. pp. 1–9 (2025) 2

    2025

  24. [24]

    arXiv preprint arXiv:2303.00915 (2023) 3, 6

    Zhang, S., Xu, Y., Usuyama, N., Xu, H., Bagga, J., Tinn, R., Preston, S., Rao, R., Wei, M., Valluri, N., et al.: Biomedclip: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs. arXiv preprint arXiv:2303.00915 (2023) 3, 6

    Pith/arXiv arXiv 2023

  25. [25]

    Advances in Neural Information Processing Systems37, 67795–67825 (2024) 2

    Zhang, T., Wang, J., Guo, H., Dai, T., Chen, B., Xia, S.T.: Boostadapter: Improv- ing vision-language test-time adaptation via regional bootstrapping. Advances in Neural Information Processing Systems37, 67795–67825 (2024) 2

    2024

  26. [26]

    Advances in Neural Information Processing Systems35, 22243–22257 (2022) 2

    Zhong, T., Chi, Z., Gu, L., Wang, Y., Yu, Y., Tang, J.: Meta-dmoe: Adapting to domain shift by meta-distillation from mixture-of-experts. Advances in Neural Information Processing Systems35, 22243–22257 (2022) 2

    2022

  27. [27]

    In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Conditional prompt learning for vision- language models. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 16816–16825 (2022) 7

    2022

  28. [28]

    International journal of computer vision130(9), 2337–2348 (2022) 7

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. International journal of computer vision130(9), 2337–2348 (2022) 7

    2022

  29. [29]

    In: Proceedings of the Computer Vision and Pattern Recognition Conference

    Zhou, L., Ye, M., Li, S., Li, N., Zhu, X., Deng, L., Liu, H., Lei, Z.: Bayesian test-time adaptation for vision-language models. In: Proceedings of the Computer Vision and Pattern Recognition Conference. pp. 29999–30009 (2025) 4

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.