{"id":"069553b0-47b9-4d83-b6e1-dfbab181db66","arxiv_id":"2608.00442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An input-conditioned, bounded prototype-pair mechanism makes CLIP-based medical anomaly detection adaptive per image and sets new state-of-the-art AUROC across six medical benchmarks.","lead":"ReCAP is a medical image anomaly detector that replaces fixed normal/abnormal reference templates with prototypes that shift slightly for each new image. It reports the top accuracy on all six zero-shot benchmarks and 23 of 24 few-shot settings, and runs over 70% faster than the fastest comparable method.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Per-element bound in Eq. (2) does not guarantee small prototype drift in high dimension; lesion contamination may still shift the normal anchor.","rationale":"The strongest empirical claim is SOTA AUROC, but the paper's unique contribution is the bounded conditioning mechanism. The most load-bearing condition for that contribution is Eq. (2)'s bound actually keeping the re-centered normal prototype close to the base prototype. I analyzed the bound dimensionally: per-coordinate bound eta does not imply small L2 displacement when C=1024. The paper's own motivation explicitly states the risk of lesion contamination, so the sufficiency of the bound is not an external assumption but the core of the method. The provided ablation (Table 5) only compares average metrics and does not measure displacement or examine large-lesion subsets. This is an internal-consistency issue, not a disagreement with consensus. If the test shows large displacements on large-lesion abnormal inputs, the central claim that bounded conditioning prevents explain-away is unverified; the method may succeed despite the mechanism. This does not overturn the empirical numbers, but it makes the paper's central explanation conditional, matching the reader's CONDITIONAL verdict. A missing baseline (UniADet) is also a concern but is less load-bearing for the mechanism's validity and can be addressed in the same revision.","tokens_in":23190,"tokens_out":7550,"duration_ms":90121,"concrete_test":"Train ReCAP using released code/checkpoints and, for all abnormal test images in BrainMRI, LiverCT, and RESC, compute the residual r in Eq. (2) and the angular displacement arccos(<q_n, ~q_n>) for the detection branch at each layer. Stratify by lesion area (e.g., top quartile of ground-truth mask size). If the mean displacement on large-lesion abnormal images exceeds, say, 10 degrees, or if the residual L2 norm is substantially larger than the inter-domain shift between source prototypes and target normal prototypes, the per-element bound is not preventing contamination. An equivalent analytical check: maximize ||sigma(beta)*eta*tanh(Wc)|| over c on the unit sphere or over the actual feature distribution and report the worst-case angular shift; if it exceeds the separation margin enforced by Eq. (12), the 'bounded' claim is false.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central safety mechanism is the bounded gated residual in Eq. (2): the claim (Sec. 3.3) is that because c_l^b(x) is pooled over the whole image, the tanh bound and sigmoid gate 'keep the residual small,' so the normal prototype shifts to the current domain but not toward the lesion. This claim is not supported by the bound actually used. With eta=0.05 and feature dimension C (ViT-L/14: C=1024), each component of r = sigma(beta)*eta*tanh(Wc) is in [-0.05, 0.05], so the L2 norm can be as large as eta*sqrt(C) ~ 1.6, which is larger than the unit norm of the prototype. After renormalization, this can rotate the anchor by tens of degrees. Nothing in the formulation regularizes W or sigma(beta) to prevent large per-image shifts; the only empirical evidence is Table 5, which averages over all test images on two datasets. If a large lesion dominates the average-pooled descriptor, the conditioned normal anchor can move toward abnormal features, and the method will partially 'explain away' the lesion. The paper acknowledges contamination but offers no analysis of whether the bound actually bounds the angular displacement. This is the weakest load-bearing assumption: if it fails, the title mechanism does not provide the claimed safety, and the SOTA numbers could be masking a systematic failure mode on large-lesion cases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes ReCAP, a language-free medical anomaly detection framework built on a frozen CLIP visual encoder. Instead of static text prompts or fixed visual anchors, ReCAP maintains normal/abnormal visual prototypes that are re-centered for each input using an average-pooled context descriptor under a bounded gated residual (Eq. 2). In the few-shot setting, a non-parametric normal-reference memory is added, and multi-layer fusion combines prototype and memory evidence. The method is evaluated with leave-one-domain-out zero-shot and K-shot protocols on six medical benchmarks, reporting best image-level AUROC on all zero-shot and 23/24 few-shot settings, best zero-shot pixel-level AUROC on all three segmentation datasets, and a >70% inference latency reduction relative to VisualAD.","tokens_in":23586,"tokens_out":5013,"duration_ms":59005,"significance":"The core idea—amortized, input-conditioned prototype re-centering without text prompts or test-time gradients—is timely and clearly relevant for cross-domain medical anomaly detection. The paper has real strengths: the method is specified precisely, the leave-one-domain-out protocol is appropriate, baselines use identical support indices, few-shot results are supplemented with mean±std over five seeds in Appendix E, and the efficiency comparison is concrete. If the bounded-modulation safety claim holds, this is a practical and competitive alternative to prompt-based and test-time adaptation methods. However, the central 'bounded drift' guarantee is not established as stated: the per-coordinate tanh bound does not by itself control the L2 norm of the residual in the 1024-dimensional feature space. In addition, hyperparameter defaults are selected from sensitivity curves computed on the same evaluation benchmarks, and zero-shot results are reported without variance. These issues do not invalidate the empirical findings, but they need to be addressed before the safety and SOTA claims can be fully accepted.","major_comments":[{"comment":"The claim that the residual is 'small' and that the conditioned anchor 'stays close to the learned base prototype' is not a consequence of the per-coordinate tanh bound. With η=0.05 and C=1024, each component of r = σ(β)η tanh(Wc) lies in [-0.05, 0.05], so the L2 norm can be as large as η√C ≈ 1.6, which exceeds the unit norm of the prototype. After renormalization, the anchor can rotate by tens of degrees. Nothing in the formulation regularizes W or σ(β) to prevent such large per-image shifts. The paper correctly acknowledges that the pooled context descriptor is contaminated by the lesion on abnormal inputs, but Table 5 reports only average AUROC on two datasets and does not measure angular drift on lesion-dominated examples. To support the central safety mechanism, the authors should either add a norm/spectral constraint that actually bounds the drift, or provide an empirical analysis","section":"§3.3, Eq. (2)"},{"comment":"The default hyperparameters η=0.05, top-k=10%, and λ=0.5 are selected from sensitivity curves computed on the same benchmark evaluation sets used to report the main results in Tables 1 and 2. The reported AUROC values are therefore not fully blind to these choices. This is especially relevant because some improvements over baselines are small (e.g., zero-shot BrainMRI 80.24 vs. 79.80 for MVFA, Table 2). I ask the authors to either tune hyperparameters on held-out source-domain data, or explicitly disclose the selection procedure and quantify its effect on the reported numbers (e.g., by reporting performance over a reasonable range of each hyperparameter for all datasets). This is not fatal to the method, but it is load-bearing for the SOTA claim.","section":"Appendix G, Figs. 6–8"},{"comment":"Zero-shot results are reported as single AUROC values without variance or number of seeds, even though training is stochastic and the improvements over the second-best method are sometimes below one percentage point (e.g., HIS 79.32 vs. 77.32 for MadCLIP; BrainMRI 80.24 vs. 79.80 for MVFA). Appendix E provides mean±std over five seeds for few-shot settings, but no analogous information is provided for the zero-shot setting that supports the headline 'best on all six datasets' claim. Please add mean±std over multiple seeds (or an explicit statement of determinism) for Table 2.","section":"Table 2 (§4.2)"}],"minor_comments":[{"comment":"The header row 'N/A ✗ ✗ ✗ ✗ ✓ ✓' is cryptic. Define the symbols in the caption or main text, including what N/A denotes for DRA.","section":"Tables 1–2 captions"},{"comment":"The caption text 'OurMVFA GTMadCLIPApril-GAN VisualAD' appears to have missing spacing/formatting. Please fix the caption formatting.","section":"Figure 2 caption"},{"comment":"The text says the fixed support set is 'selected from the validation split,' but it is not clear whether this is the same as the labeled pool described in Appendix C and whether the test set is completely disjoint. Clarify the split terminology to avoid leakage ambiguity.","section":"Appendix D, Support Sampling"},{"comment":"The normalized memory distance is mapped to [0,1] via division by 2. Since nearest-neighbor cosine distance on ℓ2-normalized features is theoretically in [0,2], this is fine, but a brief justification for using the maximum possible distance (rather than the observed maximum) would help readers.","section":"Eq. (6)"}],"recommendation":"major_revision","confidential_remarks":"The paper has a strong empirical program and a sensible evaluation protocol, but the central 'bounded' mechanism is currently more of a design heuristic than a proven safety guarantee. I would be comfortable with acceptance after the authors either impose a norm-level bound and prove the drift claim, or reframe the claim as empirical and add the contamination-stratified analysis. Also, the paper relies on several ArXiv-only 2026 references (VisualAD, UniADet); the editor may want to verify their peer-review status. No code release is mentioned, which is a limitation for reproducibility in this area."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The central mechanism—input-conditioned, bounded gated re-centering of normal/abnormal prototypes—is genuinely new for language-free medical anomaly detection, and the ablation in Table 5 shows it earns its keep against static and unbounded alternatives. That is a real method-level contribution. The second thing: the paper's headline safety claim is over-stated. The per-element bound in Eq. (2) does not control the norm of the residual: with eta=0.05 and a 1024-dim feature space, the residual's L2 norm can be ~1.6, larger than the unit-norm prototype it is shifting. So the 'bounded' design does not, by itself, guarantee small angular drift, and the acknowledged lesion-contamination risk is not actually resolved by the math. The only evidence that the mechanism stays safe is the average ablation, not an analysis of per-image drift on large-lesion cases. That is a genuine soft spot, though not fatal—the empirical pattern is consistent with the mechanism working on average.\n\nThe evaluation has real strengths. The leave-one-domain-out protocol is appropriate, support indices are shared across methods, and Appendix E gives five-seed means and stds for all few-shot settings. The speedup over VisualAD (3.5–3.8x) is credible and tied to a concrete architectural difference. The soft spots beyond the bound: zero-shot numbers in Table 2 have no variance; hyperparameters (eta, top-k, lambda) were selected using sensitivity curves on the same evaluation datasets, which is selection-on-test; and UniADet, the closest language-free baseline, is cited but omitted from all comparisons. That omission weakens the SOTA claim.\n\nBottom line: the paper deserves a serious referee. It is a solid method paper with a novel idea and a mostly fair protocol, but the central safety claim needs either an actual drift bound or a per-image empirical check. Send it to review, and ask the authors for those additions, plus zero-shot variance and a UniADet comparison.","headline":"Genuinely new conditioning mechanism with strong results, but the 'bounded' claim isn't actually bounded and the closest baseline is missing; referee it.","tokens_in":24013,"tokens_out":4385,"would_cite":false,"duration_ms":46385,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"ReCAP replaces static normal/abnormal references with per-image bounded visual prototypes, reporting best zero-shot image-level AUROC on all six medical benchmarks and over 70% lower inference latency than the fastest baseline.","keywords":["medical anomaly detection","zero-shot detection","few-shot detection","CLIP visual prototypes","bounded gated modulation","cross-domain transfer","anomaly localization","language-free"],"falsifier":"Take an abnormal image from BrainMRI or LiverCT, compute ReCAP's anomaly score, then re-run the same image after removing or heavily masking the annotated lesion region; if the score decreases (the image with the lesion looks more normal than the image without it), the conditioning is being driven by lesion content rather than domain context, falsifying the bounded-modulation premise. Repeating this across lesions of increasing size should show whether η=0.05 keeps the normal anchor clean.","tokens_in":23120,"feed_emoji":"🩺","tokens_out":5638,"duration_ms":45963,"temperature":0.7,"pith_summary":"This paper tries to establish that the standard CLIP-based anomaly detector design—a fixed 'normal' and 'abnormal' reference learned once and reused for every test image—is a liability when the detector must transfer across medical domains such as brain MRI, liver CT, and retinal OCT. It proposes ReCAP, which instead re-centers the normal/abnormal prototype pair for each input image using a bounded, gated modulation driven by that image's global appearance. The claim is that this per-image re-centering improves zero-shot image-level AUROC on all six benchmarks and most few-shot settings, while also improving pixel-level lesion segmentation and cutting inference latency by more than 70% relative to the fastest language-free baseline. A sympathetic reader would care because the method removes the need for text prompts, avoids test-time gradient updates, and offers a single forward-pass adaptation mechanism for unseen medical domains.","feed_headline":"Per-image prototypes top six medical anomaly benchmarks","feed_subtitle":"ReCAP re-centers normal/abnormal anchors per query, improving AUROC and cutting inference time by 70% without text prompts.","key_machinery":"Bounded gated prototype modulation (Eq. 2): each normal/abnormal prototype q is re-centered by σ(β) η tanh(W c(x)), where c(x) is the average-pooled, normalized patch-token descriptor of the query and η=0.05 caps the shift. The sigmoid gate and tanh bound are the load-bearing design: they let the anchor drift toward the current domain while preventing the lesion in the pooled descriptor from dragging the normal anchor onto pathological content. Around it sit the prototype separation regularizer (keeping base prototypes discriminative before adaptation), the non-parametric normal-reference memory for few-shot instance-level normality, and learnable multi-layer fusion combining prototype logit","core_discovery":"ReCAP's central claim is that an anomaly boundary expressed as a pair of visual prototypes—one normal, one abnormal—can and should be made query-adaptive, and that the adaptation must be deliberately bounded. The paper derives this from three limitations of static anchors: text prompts are coarse and wording-sensitive; learned visual tokens can still define a boundary misplaced for an unseen target; and a compact global anchor cannot preserve instance-level normal variation needed for localization. ReCAP computes a context descriptor by average-pooling the query's patch tokens, then shifts each prototype by a gated residual σ(β)η tanh(W c(x)), with η=0.05, so the conditioned prototypes stay","pith_inferences":["Editorial inference: the bounded-gate premise implies a testable monotonicity property: for a fixed image, removing or shrinking the lesion should not lower the anomaly score; if it does, the conditioning is absorbing lesion signal rather than domain signal.","Editorial inference: the method could be extended to volumetric data by treating 3D patches or slices as tokens; the context descriptor would then pool over a volume, increasing the chance of lesion contamination and requiring a smaller η or a localization-aware gate.","Editorial inference: because the approach is language-free and CLIP-based, it may transfer to industrial anomaly detection where textual prompts are also awkward, though the medical cross-domain setup (six datasets) is the paper's actual evidence.","Editorial inference: the paper's zero-shot leave-one-out protocol trains on five medical domains; a stronger test would be holding out a modality never seen at all (e.g., ultrasound), which the paper lists as future direction."],"forward_implications":["If ReCAP's central claim holds, language-free, gradient-free per-image anchoring becomes a competitive alternative to text-prompt CLIP adaptation for medical anomaly detection.","The bounded conditioning design implies that a single source-trained prototype pair can be shipped to unseen target domains and adapted in one forward pass at test time.","The normal-reference memory makes few-shot detection benefit from instance-level normal variation rather than only a compact prototype; ablations attribute gains to memory, conditioning, and layer fusion jointly.","The reported ~3.8x latency reduction over VisualAD on the same ViT-L/14@336px backbone suggests the method is suitable for time-sensitive screening if the accuracy results replicate.","The sensitivity analyses (η around 0.025–0.05, top-k ratio 5–20%, fusion weight λ around 0.5) indicate the method's gains do not hinge on a narrow hyperparameter setting."],"supporting_citations":[{"why":"Provides the frozen CLIP visual encoder whose feature space all prototypes and conditioning operate in.","marker":"Radford et al. 2021"},{"why":"VisualAD, the language-free static-anchor baseline that ReCAP compares against and the primary speed comparison (66.1 ms vs 17.4 ms).","marker":"Hou et al. 2026"},{"why":"UniADet, the other language-free CLIP-based baseline with task- and level-specific visual decision weights.","marker":"Gao and Wang 2026"},{"why":"MVFA, a multi-level visual adaptation baseline and the benchmark protocol that defines the six medical datasets and leave-one-out evaluation.","marker":"Huang et al. 2024"},{"why":"MadCLIP, the few-shot CLIP baseline that ReCAP outperforms on most settings and the strongest prompt-based competitor.","marker":"Shiri, Beyan, and Murino 2025"},{"why":"MediCLIP, a few-shot medical CLIP adaptation baseline that frames the text-prompt approach ReCAP abandons.","marker":"Zhang et al. 2024"}],"fun_headline_variants":["Query-adaptive prototypes beat static anchors across six benchmarks","ReCAP: adaptive per-image prototypes slash inference latency by 70%","Language-free ReCAP uses bounded prototype conditioning to top six benchmarks","Bounded prototype conditioning: query-adaptive anchors top six benchmarks","ReCAP re-centers anomaly prototypes for each image"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"Equation (2) presumes that the average-pooled context descriptor carries domain information without being so dominated by a large lesion that the bounded gate shifts the normal prototype toward the anomaly; the paper does not analyze this contamination failure mode.","fun_headline_variants_meta":{"raw":{"variants":["Query-adaptive prototypes beat static anchors across six benchmarks","ReCAP: adaptive per-image prototypes slash inference latency by 70%","Language-free ReCAP uses bounded prototype conditioning to top six benchmarks","Bounded prototype conditioning: query-adaptive anchors top six benchmarks","ReCAP re-centers anomaly prototypes for each image"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000922,"raw_usage":{"total_tokens":3787,"prompt_tokens":734,"completion_tokens":3053,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":2969}},"tokens_in":478,"tokens_out":3053,"duration_ms":16726,"temperature":1.0,"reasoning_tokens":2969,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T00:57:24.950708+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an abnormal image from BrainMRI or LiverCT, compute ReCAP's anomaly score, then re-run the same image after removing or heavily masking the annotated lesion region; if the score decreases (the image with the lesion looks more normal than the image without it), the conditioning is being driven by lesion content rather than domain context, falsifying the bounded-modulation premise. Repeating this across lesions of increasing size should show whether η=0.05 keeps the normal anchor clean.","supporting_citations":[],"review_version":1}