{"id":"94fd03b5-9764-4476-9732-69815a4c5b0d","arxiv_id":"2606.07102","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":1,"one_line_summary":"GP-Adapter builds modality-specific one-class GPs on CLIP image and text features using RBF and linear kernels to generate variance-aware scores that improve few-shot OOD detection.","lead":"GP-Adapter adds class-wise Gaussian Processes on frozen CLIP embeddings to produce uncertainty estimates for few-shot out-of-distribution detection without any model training. A smart generalist might read it to see a lightweight way to make large vision-language models more reliable when data is scarce and inputs shift.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Whether fused class-wise GP predictive variances reliably separate ID from OOD samples remains the central unverified assumption.","rationale":"The reader's weakest assumption directly identifies the empirical linchpin of the OOD claim. Because the manuscript supplies only experimental tables rather than an ablation isolating the fusion rule or a sensitivity analysis on kernel hyperparameters, the same concern remains load-bearing even after reading the full text. No independent verification (e.g., code release with exact hyperparameter logs) alters this assessment, so the reader's UNVERDICTED stance is retained.","tokens_in":1736,"tokens_out":409,"duration_ms":20850,"concrete_test":"Recompute the OOD AUROC tables using (a) image-GP variance alone, (b) text-GP variance alone, and (c) the paper's fused score on the same ImageNet OOD splits; if the fused score does not exceed the better of (a) or (b) by a statistically significant margin on at least three benchmarks, the claimed complementarity of the GP term is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline experimental claim (competitive few-shot ID accuracy plus improved OOD detection via complementarity with prompt learning) rests on the fused variance score being higher for OOD inputs than ID inputs across the reported benchmarks. The construction uses one-class GPs per class (RBF kernel on image embeddings, linear kernel on text prompts) whose predictive variances are combined into a single confidence score. In the few-shot regime the RBF length-scale and noise variance are the only tunable parameters; if these are chosen without an explicit ID-only validation protocol, or if the fusion rule (max, product, or learned) does not provably emphasize distance from all class-conditional GPs, the observed OOD gains could be driven by the prompt-learning baseline rather than the GP uncertainty term. No section of the manuscript supplies a derivation showing that the fused variance is monotonic with distributional distance in CLIP space.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes GP-Adapter, a training-free framework augmenting frozen CLIP with class-wise one-class Gaussian Processes (RBF kernel on image embeddings, linear kernel on text prompts) whose predictive variances are fused into a variance-aware confidence score for few-shot classification and OOD detection. It reports competitive few-shot accuracy on ImageNet and consistent OOD gains when combined with prompt-learning baselines, with memory scaling O(CK^2) and code released.","tokens_in":1928,"tokens_out":468,"duration_ms":18431,"significance":"If the central empirical claim holds, the work supplies a lightweight, non-fine-tuning route to uncertainty quantification on CLIP embeddings that complements prompt learning, with explicit reproducibility via code release and a clear complexity bound. This could be useful for low-data reliability under distribution shift.","major_comments":[{"comment":"The central OOD claim rests on the fused class-wise GP predictive variance being reliably higher for OOD than ID inputs, yet no derivation or monotonicity argument is supplied showing that the fusion (max/product/learned) of one-class GP variances separates distributional distance in CLIP space; this is load-bearing for attributing gains to the GP term rather than the prompt baseline.","section":"Method"},{"comment":"The hyperparameter selection for the RBF length-scale and noise variance is described only as 'lightweight'; without an explicit ID-only validation protocol stated, it is unclear whether the reported OOD improvements could be driven by implicit leakage or baseline effects rather than the uncertainty model.","section":"Experiments"}],"minor_comments":[{"comment":"The abstract asserts 'competitive few-shot performance' and 'consistent improvements' but supplies no numerical values, tables, or error bars; these should be summarized with at least one key metric (e.g., AUROC or accuracy delta) for immediate readability.","section":"Abstract"},{"comment":"Notation for the fusion rule and the exact form of the variance-aware score is introduced without an equation number; adding an explicit equation would clarify how image and text GP outputs are combined.","section":"Method"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments and the opportunity to clarify our work. Below we respond point-by-point to the major comments and indicate the revisions we will make.","responses":[{"response":"We agree that a formal derivation of monotonicity for the variance fusion is absent. GP-Adapter is an empirical framework that builds on the standard property of one-class GPs: predictive variance increases with distance from the inducing points in the embedding space. The class-wise one-class GPs (RBF on images, linear on text) are intended to capture this effect per class, and the fusion (max, product, or learned) is chosen to produce a scalar uncertainty signal. While we do not claim a theoretical guarantee that the fused variance strictly separates ID from OOD in CLIP space, the reported experiments show consistent OOD AUROC/AUPR gains when the GP term is added to prompt-learning baselines. In revision we will expand the method section with additional intuition, a small illustrative 2-D example of variance behavior, and explicit caveats that the separation is demonstrated empirically rather than proven. We will also emphasize that improvements are measured relative to the same prompt baseline, so the incremental benefit is attributable to the added GP variance.","revision_made":"partial","referee_comment":"[Method] The central OOD claim rests on the fused class-wise GP predictive variance being reliably higher for OOD than ID inputs, yet no derivation or monotonicity argument is supplied showing that the fusion (max/product/learned) of one-class GP variances separates distributional distance in CLIP space; this is load-bearing for attributing gains to the GP term rather than the prompt baseline."},{"response":"We thank the referee for highlighting this ambiguity. The length-scale and noise variance are selected via a small held-out subset of the ID few-shot examples (typically 20 % of the K-shot cache per class) using a simple validation criterion on ID classification accuracy; no OOD data is ever used. We will revise the experimental setup and hyperparameter section to state this ID-only protocol explicitly, including the exact split and selection criterion, so that readers can verify the absence of leakage.","revision_made":"yes","referee_comment":"[Experiments] The hyperparameter selection for the RBF length-scale and noise variance is described only as 'lightweight'; without an explicit ID-only validation protocol stated, it is unclear whether the reported OOD improvements could be driven by implicit leakage or baseline effects rather than the uncertainty model."}],"tokens_in":1365,"tokens_out":499,"duration_ms":17491,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The paper's core move is to fit class-wise one-class GPs on top of frozen CLIP features, using RBF on the image side and linear on the text side, then combine the predictive variances into an OOD score. It stays training-free, keeps memory at O(C K squared), and releases code. That setup is the concrete new piece: a lightweight probabilistic wrapper that can sit next to prompt-learning methods without touching the backbone.\n\nThe experiments are described as showing competitive few-shot ImageNet accuracy plus consistent OOD gains when the GP term is added to prompt baselines. If those tables hold up, the complementarity claim is useful for anyone already running prompt tuning in low-data regimes.\n\nThe main gap is that the abstract supplies no numbers, no fusion-rule ablations, and no protocol for choosing the RBF length-scale and noise variance. The stress-test note is right on this: without an explicit ID-only validation split or a derivation that the fused variance grows with distributional distance, it is possible the reported OOD lift is carried by the prompt baseline rather than the GP term. The central assumption—that the variance reliably flags OOD across the tested benchmarks—still needs the full results to check.\n\nThis is for people who need a plug-in uncertainty layer for CLIP-style models under data scarcity and shift. A reader already working on reliable vision-language pipelines would find the construction straightforward to try. It deserves a serious referee because the method is simple, the code is public, and the empirical claims are falsifiable once the tables are examined.","headline":"GP-Adapter adds modality-specific one-class GPs on frozen CLIP embeddings for a training-free variance score in few-shot OOD, but the abstract leaves the actual separation quality and ablation results unshown.","tokens_in":2405,"tokens_out":397,"would_cite":false,"duration_ms":14983,"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":"GP-Adapter adds class-wise Gaussian processes to frozen CLIP embeddings to produce variance-aware scores for few-shot out-of-distribution detection.","keywords":["few-shot classification","out-of-distribution detection","Gaussian processes","CLIP adapter","uncertainty estimation","vision-language models","training-free adaptation"],"falsifier":"An OOD benchmark in which the AUROC or FPR95 of the fused GP variance score is no better than, or worse than, the deterministic CLIP similarity baseline.","tokens_in":2645,"feed_emoji":"📊","tokens_out":620,"duration_ms":13203,"temperature":0.7,"pith_summary":"The paper introduces a training-free method that places one-class Gaussian processes on top of CLIP image and text embeddings to supply uncertainty estimates where standard similarity scores provide none. Separate GPs are fit per class using an RBF kernel on images and a linear kernel on prompts, then their predictive variances are fused into a single confidence measure. This construction requires only a small cache of K-shot examples per class and no updates to the underlying CLIP weights. A reader would care because the resulting variance signal improves detection of samples from unseen distributions while remaining competitive on standard few-shot classification tasks. The experiments further indicate that the uncertainty signal complements existing prompt-tuning techniques rather than replacing them.","feed_headline":"Gaussian processes on CLIP embeddings sharpen few-shot OOD detection","feed_subtitle":"Training-free adapter fuses image and text predictive variances to flag distribution shifts without updating the backbone","key_machinery":"Modality-specific class-wise one-class Gaussian processes whose predictive variances are fused into a single OOD confidence score.","core_discovery":"GP-Adapter constructs modality-specific, class-wise one-class GPs on top of frozen CLIP embeddings using an RBF kernel for image features and a linear kernel for text prompts and fuses their predictive statistics to produce a variance-aware confidence score for OOD detection.","pith_inferences":["The same variance-fusion idea could be applied to other frozen vision-language backbones by swapping only the embedding extractor.","In safety-critical low-data settings the explicit uncertainty signal may reduce over-confident errors on shifted inputs.","Replacing the fixed RBF and linear kernels with a learned or adaptive kernel choice might further tighten the separation between in- and out-of-distribution variances."],"forward_implications":["Competitive accuracy on ImageNet few-shot classification using only the cached examples.","Consistent gains in OOD detection metrics when the GP variance is added to prompt-learning baselines.","Memory requirement scales as O(C K squared) where C is the number of classes and K the shots per class.","The method works without any gradient updates to the CLIP backbone."],"fun_headline_variants":["Gaussian processes applied to CLIP embeddings for few-shot OOD","Class-wise one-class GPs on CLIP with RBF kernel for image features","Fusing predictive variances from image and text GPs on frozen CLIP","Training-free GP-Adapter produces variance-aware OOD confidence scores","Modality-specific GPs on CLIP embeddings for OOD detection"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The fused predictive variance from the class-wise GPs reliably separates in-distribution examples from out-of-distribution examples on the tested benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian processes applied to CLIP embeddings for few-shot OOD","Class-wise one-class GPs on CLIP with RBF kernel for image features","Fusing predictive variances from image and text GPs on frozen CLIP","Training-free GP-Adapter produces variance-aware OOD confidence scores","Modality-specific GPs on CLIP embeddings for OOD detection"]},"model":"grok-4.3","cost_usd":0.006325,"raw_usage":{"total_tokens":2970,"prompt_tokens":665,"num_sources_used":0,"completion_tokens":87,"cost_in_usd_ticks":63249500,"prompt_tokens_details":{"text_tokens":665,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2218,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":665,"tokens_out":87,"duration_ms":14335,"temperature":1.0,"reasoning_tokens":2218,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-27T22:21:00.287773+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"An OOD benchmark in which the AUROC or FPR95 of the fused GP variance score is no better than, or worse than, the deterministic CLIP similarity baseline.","supporting_citations":[],"review_version":1}