{"id":"88fe1097-9246-4a6d-9e54-de068612e241","arxiv_id":"2607.16726","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"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.","lead":"The paper introduces MoBE, a no-gradient method that makes medical vision-language models pick and adjust among modality experts during inference. It reports average accuracy gains of 4.3 to 7.2 points across seen, unseen, and heterogeneous medical benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline gains rest on unreported per-dataset hyperparameters and single runs; without exact values and error bars the central claim is not testable.","rationale":"The reader's EBA error-accumulation concern is real and worth testing, but it is a robustness caveat, not the first blocker for the central empirical claim. The first blocker is that the reader already notes unreported hyperparameters, yet lists them as reporting gaps rather than the weakest assumption. I consider the missing per-dataset configuration to be more load-bearing: it directly determines whether the +4.72/+7.17/+4.3 numbers can be reproduced and whether MoBE beats strong baselines fairly. Because the reader's verdict is already CONDITIONAL for these reasons, my analysis does not move the verdict; it sharpens the reason. I give partial agreement because the reader identified part of the same issue (reporting gaps) but chose the EBA loop as the weakest assumption.","tokens_in":8644,"tokens_out":6608,"duration_ms":68788,"concrete_test":"Obtain the exact per-dataset hyperparameters and runnable code from the authors. Then rerun Tables 1–3 under two conditions: (a) the reported per-dataset values, and (b) a single fixed configuration (e.g., τ=0.5, τ_EBA=1.0, θ1=0.8, θ2=0.9, λ=0.5) with the same expert bank. For each condition, run at least 5 random seeds/stream orders and report mean±std. If the fixed-config gains fall below the claimed +4.72/+7.17/+4.3 or if the per-dataset vs fixed gap exceeds ~2 points, the headline result is dependent on per-dataset tuning and the central claim needs to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MoBE's training-free routing plus Bayesian online statistics yields +4.72/+7.17/+4.3 over SOTA TTA. For this to hold, the reported numbers must reflect the method, not per-dataset tuning. §3.1 states 'Dataset-specific hyperparameter values for MoBE are available at Github' but no URL or values appear. MoBE has at least five configurable quantities per dataset: entropy gap τ (Eq. 1), inverse-entropy weighting (Eq. 2), EBA scale τ_EBA, confidence thresholds θ1/θ2 (Eqs. 4–5), and fusion weight λ. All tables report a single run, with no error bars or seeds. Ablations in §3.3 are also single-run and presumably use the same per-dataset values. If τ/θ1/θ2/λ were selected on each test set, the comparison to fixed-hyperparameter baselines (TENT, TDA, MoME) is not apples-to-apples and the central claim is not established. Supporting reporting issues: Table 4's seen/unseen averages (35.45/23.01 for BiomedCLIP) do not match Tables 1–2 (33.00/20.49), and no aggregation definition is given.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8958,"tokens_out":3033,"duration_ms":32010,"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":[{"comment":"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.","section":"§3.1"},{"comment":"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.","section":"Table 4"},{"comment":"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.","section":"§2.2, Eqs. (4)–(5)"}],"minor_comments":[{"comment":"The code/dataset link is given as a bare 'Github' placeholder; an actual URL is needed.","section":"Abstract/§3.1"},{"comment":"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.","section":"Fig. 1b"},{"comment":"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.","section":"Tables 1–2"},{"comment":"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.","section":"§3.1"},{"comment":"The abbreviation 'MOME' appears in Table 2 while 'MoME' is used elsewhere; please standardize the capitalization.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the benchmark setup is relevant, but the missing hyperparameter report and the Table 4 inconsistency prevent the central empirical claim from being verified. These issues are fixable within the manuscript's scope, so I recommend major revision rather than rejection. I would also ask the authors to explicitly state how hyperparameters were chosen and to provide runs across seeds or a small validation-based selection protocol, as this is essential for a fair comparison in TTA."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nHere is my quick read of arXiv:2607.16726. The core idea — dynamic-k entropy routing over modality experts plus per-expert Bayesian prototype/prior adaptation, all without backprop — is genuinely new and clearly presented. It doesn't just bolt entropy routing onto MoME; the confidence-gated updates of both prototypes and class priors are a meaningful twist. The ablations support each component: dynamic-k beats fixed top-k, inverse-entropy weighting helps, and EBA contributes the largest gain. That is honest, incremental engineering work.\n\nThe problem is in the reporting. The method has at least five per-dataset hyperparameters (tau, theta1, theta2, lambda, tau_EBA), and the paper says values are \"available at Github\" but gives no URL or commit hash. Without those, the reader cannot reproduce the numbers, and — more importantly — the comparison against fixed-hyperparameter baselines like TENT, MoME, and TDA is not apples-to-apples. If these were tuned on each test set, the headline +4.72/+7.17/+4.3 gains are overstated. Single-run results with no error bars or seeds deepen the concern. Table 4's averages for BiomedCLIP (35.45/23.01) also don't match Tables 1–2 (33.00/20.49), and the aggregation is undefined. That is a small fix, but it signals carelessness.\n\nThere is also a real soft spot in the EBA loop. If an expert is confidently wrong under an unseen-modality shift, it updates its own prototype and prior toward the wrong class, and the corruption can compound over the stream. The paper doesn't analyze this or give evidence that the confidence thresholds bound the damage. It may be fine in practice — the method still improves averages — but it deserves discussion.\n\nWho gets value from this? Researchers working on inference-only TTA for medical VLMs, and anyone who wants a concrete alternative to gradient-based MoE adaptation. As it stands, the experimental claims are not fully testable.\n\nMy recommendation: send it to peer review, but with the expectation of major revision on transparency. If the authors release the exact per-dataset hyperparameters and multiple seeds, and the gains hold, this is a useful contribution. If the hyperparameters were tuned per test set, the comparison needs to be reframed.\n\nBest,\n[Your name]","headline":"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.","tokens_in":9449,"tokens_out":4235,"would_cite":true,"duration_ms":40528,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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","keywords":["test-time adaptation","mixture of experts","medical vision-language models","modality generalization","entropy routing","Bayesian adaptation","BiomedCLIP","optimization-free inference"],"falsifier":"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.","tokens_in":8543,"feed_emoji":"🩺","tokens_out":4703,"duration_ms":41695,"temperature":0.7,"pith_summary":"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.","feed_headline":"No-training expert routing gains 4–7 points on unseen medical modalities","feed_subtitle":"Entropy picks the right modality experts; online Bayesian statistics refine them without gradient updates.","key_machinery":"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","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"fun_headline_variants":["Frozen MVLMs adapt to new modalities via test-time routing, no training","Optimization-free expert routing boosts medical modality accuracy by up to 7 points","Test-time expert selection without gradients gains 4–7 on medical benchmarks","MoBE: training-free test-time routing adapts MVLMs to unseen modalities","Entropy-guided routing tunes medical MVLMs at test time, no training needed"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Frozen MVLMs adapt to new modalities via test-time routing, no training","Optimization-free expert routing boosts medical modality accuracy by up to 7 points","Test-time expert selection without gradients gains 4–7 on medical benchmarks","MoBE: training-free test-time routing adapts MVLMs to unseen modalities","Entropy-guided routing tunes medical MVLMs at test time, no training needed"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001607,"raw_usage":{"total_tokens":6222,"prompt_tokens":716,"completion_tokens":5506,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":5403}},"tokens_in":460,"tokens_out":5506,"duration_ms":29459,"temperature":1.0,"reasoning_tokens":5403,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T20:06:30.543625+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[],"review_version":1}