{"id":"c15915fc-9edc-40a7-ad6b-3b09526e7f12","arxiv_id":"2606.12925","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":7.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"BCP refines zero-shot logits via anchor-conditioned Bayesian updates estimated online from test-stream co-occurrences to promote compatible labels and improve mAP in multi-label classification.","lead":"The paper proposes Bayesian Conditional Priors (BCP) as a gradient-free test-time adaptation technique for multi-label recognition using frozen vision-language models under distribution shift. It estimates label co-occurrence structure from unlabeled test data to refine zero-shot predictions without tuning the backbone.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Zero-shot logits as proxies for marginal posteriors under fixed likelihood may not isolate prior mismatch as the dominant error source","rationale":"The reader's weakest_assumption directly identifies the modeling assumption whose violation would invalidate the entire Bayesian refinement pipeline; no other internal inconsistency is visible from the provided abstract and claim. Because the review was abstract-only, the concrete_test above is the minimal verification step that would either confirm or refute the separation before any performance numbers can be trusted.","tokens_in":1726,"tokens_out":355,"duration_ms":8450,"concrete_test":"On one benchmark (e.g., the RN50 COCO-to-VOC shift), compute the zero-shot logits on both source and target images; fit a supervised logistic regression on source labels to obtain an empirical marginal posterior; measure the KL divergence between the zero-shot proxy and the supervised posterior on target images. If the divergence is large and the BCP gain disappears when the same co-occurrence statistics are applied to the supervised posterior instead of the zero-shot logits, the proxy assumption is the dominant failure mode.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central construction treats zero-shot logits as p(y|x) under an unchanging image-text likelihood p(x|y), so that any performance drop is ascribed to a mismatched label prior p(y) that BCP then corrects via anchor-conditioned online co-occurrence statistics. This decomposition is load-bearing: if the VLM likelihood itself changes under the target distribution (e.g., via altered visual features or text embeddings), the closed-form logit-space update and PMI interpretation no longer separate prior from likelihood errors. The abstract and method description provide no diagnostic that would falsify this separation, nor any comparison against a pure likelihood-adaptation baseline.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces Bayesian Conditional Priors (BCP) Estimation, a gradient-free test-time adaptation technique for multi-label recognition with frozen VLMs such as CLIP. It treats zero-shot logits as proxies for marginal posteriors under a fixed image-text likelihood, attributes distribution-shift errors primarily to mismatched label priors, and corrects them via an anchor-conditioned Bayesian update in logit space that admits a PMI interpretation. Anchor-conditioned co-occurrence statistics are estimated online from the unlabeled test stream using lightweight second-order moments, enabling closed-form refinement without target labels or backbone tuning. Experiments across multi-label benchmarks report consistent gains, e.g., RN50 mAP rising from 57.31 to 69.22 and ViT-B/16 from 62.61 to 71.79 over strong TTA baselines.","tokens_in":1880,"tokens_out":551,"duration_ms":14945,"significance":"If the central decomposition and online estimation hold, BCP supplies an efficient, annotation-free route to inject label co-occurrence structure into zero-shot multi-label inference, with negligible overhead beyond one forward pass. The closed-form logit-space update and reproducible online statistics constitute concrete strengths that could influence practical VLM deployment under label-distribution shift.","major_comments":[{"comment":"§3 (Method), the load-bearing assumption that zero-shot logits equal marginal posteriors p(y|x) under an unchanging likelihood p(x|y): the manuscript provides no diagnostic experiment (e.g., controlled likelihood-shift simulation or comparison against a pure likelihood-adaptation baseline) that would falsify the separation of prior versus likelihood error sources. Without such a test the attribution of gains to prior correction remains unverified.","section":"§3"},{"comment":"§3.3 (online estimation), the precise procedure for computing anchor-conditioned co-occurrence statistics from the streaming test batch: the description is high-level and the independence from any training-set quantities is asserted but not shown via an explicit equation or pseudocode that would allow verification that the update remains purely test-time.","section":"§3.3"}],"minor_comments":[{"comment":"Table 2 and 3: report per-dataset standard deviations or number of runs so that the magnitude of the reported mAP gains can be assessed for statistical reliability.","section":"Tables 2,3"},{"comment":"§4.2: the anchor-selection rule (high-confidence threshold or top-k) is stated qualitatively; an explicit formula or hyper-parameter sensitivity plot would improve reproducibility.","section":"§4.2"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive feedback. We address each major comment below and will revise the manuscript accordingly to strengthen the presentation of our assumptions and implementation details.","responses":[{"response":"We agree that the central modeling choice—treating zero-shot logits as proxies for marginal posteriors under a fixed likelihood while attributing shift errors primarily to mismatched priors—would benefit from explicit verification. The manuscript motivates this decomposition via the Bayesian update and PMI interpretation, and the empirical gains over strong TTA baselines are consistent with prior correction, yet we acknowledge the absence of a controlled diagnostic. In the revision we will add a simulation experiment that artificially perturbs the likelihood (e.g., via synthetic visual-feature shifts) while holding the label prior fixed, together with a comparison against a pure likelihood-adaptation baseline, to better isolate the contribution of the prior update.","revision_made":"yes","referee_comment":"[§3] §3 (Method), the load-bearing assumption that zero-shot logits equal marginal posteriors p(y|x) under an unchanging likelihood p(x|y): the manuscript provides no diagnostic experiment (e.g., controlled likelihood-shift simulation or comparison against a pure likelihood-adaptation baseline) that would falsify the separation of prior versus likelihood error sources. Without such a test the attribution of gains to prior correction remains unverified."},{"response":"We accept that the online estimation procedure in §3.3 is described at a high level and that explicit verification of its test-time-only nature is needed. The method relies exclusively on second-order moments accumulated from the unlabeled test stream; no training-set statistics are used. In the revision we will supply the precise update equations for the anchor-conditioned co-occurrence matrix and include pseudocode that makes the streaming, training-set-independent computation fully verifiable.","revision_made":"yes","referee_comment":"[§3.3] §3.3 (online estimation), the precise procedure for computing anchor-conditioned co-occurrence statistics from the streaming test batch: the description is high-level and the independence from any training-set quantities is asserted but not shown via an explicit equation or pseudocode that would allow verification that the update remains purely test-time."}],"tokens_in":1429,"tokens_out":469,"duration_ms":16592,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The one thing to know is that BCP adds label co-occurrence structure to zero-shot multi-label predictions at test time using a closed-form Bayesian update conditioned on an anchor label, with priors estimated online from the test stream. It claims big improvements over baselines on CLIP models.\n\nThe combination of anchor selection, logit-space update with PMI reading, and second-order co-occurrence stats from unlabeled data looks new and not just an obvious extension. It does well by staying gradient-free and adding almost no compute beyond the forward pass, which makes it deployable.\n\nThe reported results show clear lifts, for example RN50 mAP from 57.31 to 69.22. That level of gain on standard benchmarks is worth attention if it holds.\n\nThe main concern is whether the fixed-likelihood assumption actually holds. The method treats zero-shot logits as marginal posteriors and blames the prior for the shift errors, but if the distribution shift also alters how the VLM matches images to text, then the update may be correcting the wrong thing. The abstract gives no experiment that would show this separation is valid, such as a comparison to likelihood adaptation. The exact rules for choosing the anchor and computing the co-occurrences are also not spelled out, so it's difficult to assess if the procedure is fully specified or sensitive to implementation details.\n\nThis paper is aimed at researchers working on test-time adaptation and efficient use of VLMs for multi-label tasks. It is worth sending for peer review because the idea is distinct and the empirical claims are strong enough to merit checking the full methods and experiments.","headline":"BCP gives a practical gradient-free TTA method for multi-label VLMs by anchor-conditioned Bayesian prior updates from test-stream co-occurrences, with large reported gains, but the fixed-likelihood assumption that isolates prior mismatch as the main error source is untested and load-bearing.","tokens_in":2367,"tokens_out":418,"would_cite":false,"duration_ms":22325,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Bayesian Conditional Priors refine zero-shot multi-label predictions by estimating conditional label dependencies online from the unlabeled test stream.","keywords":["multi-label recognition","test-time adaptation","Bayesian conditional priors","vision-language models","distribution shift","label co-occurrence","zero-shot inference"],"falsifier":"If the Bayesian update fails to raise mAP on held-out shifted multi-label test sets relative to the strongest TTA baselines, or if the online co-occurrence estimates do not improve over independent-label baselines.","tokens_in":2642,"feed_emoji":"🔄","tokens_out":658,"duration_ms":14736,"temperature":0.7,"pith_summary":"The paper claims that errors in zero-shot multi-label inference under distribution shift arise primarily from mismatched label priors rather than the underlying image-text likelihood. It introduces a gradient-free method that selects a high-confidence anchor label per test image and applies a closed-form Bayesian update in logit space to promote labels that co-occur with the anchor. The update admits a pointwise mutual information reading and relies on lightweight second-order co-occurrence statistics collected on the fly from the test batch. No backbone tuning or target labels are required. Reported results show large mAP gains across CLIP backbones on standard multi-label benchmarks.","feed_headline":"Bayesian priors correct incoherent multi-label predictions at test time","feed_subtitle":"BCP estimates label co-occurrences from unlabeled test images and applies a closed-form update, lifting RN50 mAP from 57.31 to 69.22 without","key_machinery":"Anchor-conditioned Bayesian refinement: a closed-form logit-space update, interpretable via pointwise mutual information, that raises the posterior of labels compatible with a chosen high-confidence anchor while lowering the posterior of incompatible ones.","core_discovery":"BCP treats zero-shot logits as proxies for marginal posteriors and corrects shift-induced errors by performing an anchor-conditioned Bayesian refinement whose logit-space update explicitly incorporates label co-occurrence structure estimated from the unlabeled test stream.","pith_inferences":["The method suggests that many existing TTA techniques could be strengthened by explicitly modeling label priors instead of only adapting features or thresholds.","If test streams are short or non-stationary, the online second-order statistics may need regularization or a decay factor to remain accurate.","The PMI interpretation opens a route to replace the Bayesian update with other information-theoretic measures of label compatibility if desired."],"forward_implications":["Label predictions become more coherent because the update promotes labels that actually co-occur with the anchor.","The adaptation requires only a single forward pass plus cheap second-order statistics and therefore scales to long test streams.","Performance gains hold across multiple CLIP backbones without any gradient steps on the model.","The same prior-estimation machinery can be applied at inference time to any frozen VLM used for multi-label tasks."],"fun_headline_variants":["Bayesian Conditional Priors correct multi-label shift errors at test time","Anchor-conditioned Bayesian updates fix incoherent zero-shot label sets","BCP estimates label co-occurrences for Bayesian refinement in multi-label tasks","Closed-form logit updates incorporate priors from unlabeled test streams"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Zero-shot logits can be treated as reliable proxies for marginal posteriors under a fixed image-text likelihood, with most shift errors coming from an incorrect label prior.","fun_headline_variants_meta":{"raw":{"variants":["Bayesian Conditional Priors correct multi-label shift errors at test time","Anchor-conditioned Bayesian updates fix incoherent zero-shot label sets","BCP estimates label co-occurrences for Bayesian refinement in multi-label tasks","Closed-form logit updates incorporate priors from unlabeled test streams"]},"model":"grok-4.3","cost_usd":0.00372,"raw_usage":{"total_tokens":1916,"prompt_tokens":641,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":37199500,"prompt_tokens_details":{"text_tokens":641,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1207,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":641,"tokens_out":68,"duration_ms":7307,"temperature":1.0,"reasoning_tokens":1207,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T07:36:34.661397+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"If the Bayesian update fails to raise mAP on held-out shifted multi-label test sets relative to the strongest TTA baselines, or if the online co-occurrence estimates do not improve over independent-label baselines.","supporting_citations":[],"review_version":1}