{"id":"db3eb570-4f8a-48f2-abb0-6cd1e46fb373","arxiv_id":"2411.15736","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"GaCoOp applies the PCGrad/ProGrad gradient projection to LoCoOp's two losses and reports improved FPR95 on ImageNet few-shot OOD detection.","lead":"Few-shot OOD detection with CLIP often struggles because the gradient for classifying in-distribution images and the gradient for regularizing background 'out-of-distribution' regions can point against each other. GaCoOp adds a gradient-projection step that removes the conflicting component, and reports lower FPR95 on ImageNet OOD benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) drops the OOD regularization gradient from the update, so the claimed conflict mitigation is not actually optimizing the two-term objective; missing ablations and angle statistics leave the mechanism unsupported.","rationale":"After reading the paper in good faith, I find the update rule Eq. (4) is the heart of the method, and it is precisely a ProGrad/PCGrad-style projection with Go as the reference direction. The paper's narrative is that this rule mitigates conflict between the two terms of Eq. (3). However, inspecting Eq. (4) carefully, Go appears only in the condition and in the projection coefficient; it is never added to the update. In the aligned case, the update is just Gi; in the conflicting case, it is Gi stripped of its component along Go. Thus the OOD regularization loss Lood is not being minimized at any step; it serves only as a local constraint on the ID gradient. The central causal claim that 'GaCoOp mitigates the gradient conflict between ID classification optimization and OOD regularization' is therefore not directly established: a method that does not optimize one of the two conflicting objectives cannot be said to align them. The experimental results may still be valid—the projection could improve OOD detection by regularizing ID features—but the paper provides no ablation separating the effect of the projection from the effect of removing λGo, nor any statistic (e.g., angle distribution, loss curves) confirming that conflicts actually occur and are reduced. This is the load-bearing gap. The reader's weakest assumption—that the angle between Gi and Go fully captures conflict and that Go is a trustworthy reference—is related, but the deeper issue is that Go is not part of the optimized objective. A concrete ablation with variants (c) and (d) above would settle whether the missing Go term matters. I would keep the conditional verdict: the method is promising, but acceptance should require these clarifying experiments.","tokens_in":8088,"tokens_out":10665,"duration_ms":101084,"concrete_test":"Run four variants with the same seeds and hyperparameters: (a) LoCoOp: Gi + λGo; (b) CoOp: Gi; (c) GaCoOp: Eq. (4) with no Go term; (d) GaCoOp+Go: Eq. (4) update plus λGo. Report AUROC, FPR95, ID accuracy, and the fraction of steps with Gi·Go < 0 (or mean angle). If (c) ≈ (d), dropping Go is harmless and the projection is the active ingredient; if (d) > (c), the method is missing a useful OOD signal and the 'conflict mitigation' story is incomplete. Also report Lood curves for (a) vs (c) to verify whether GaCoOp actually reduces OOD entropy.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is that Eq. (4) does not implement optimization of the objective in Eq. (3). In both branches, the update is a function of Gi only: either Gi itself, or Gi with its component along Go removed. The OOD regularization gradient Go never enters the parameter update as an optimization direction; it is used only as a reference to decide whether to project Gi. Thus GaCoOp does not minimize Lood at all—it minimizes Lcoop along a direction that is constrained to be orthogonal to Go when the angle is obtuse (and unconstrained when acute). The paper frames the central claim as 'mitigating gradient conflict between ID classification and OOD regularization,' but removing one of the two objectives is a different mechanism. The reported gains over LoCoOp may come from the projection acting as a regularizer on the ID gradient, or from simply avoiding the potentially harmful OOD term, rather than from 'aligning' two co-optimized losses. The paper provides no ablation of the λGo term, no measurement of how often the obtuse branch triggers, and no angle distribution or Lood trajectory, so the causal claim that gradient conflict is reduced is not directly supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes GaCoOp, a prompt-tuning method for few-shot out-of-distribution (OOD) detection with CLIP. It builds on LoCoOp's two-term objective (3), which combines an ID cross-entropy classification loss with an OOD entropy regularization loss, and modifies the update rule so that at each step the ID classification gradient Gi is either used directly when Gi·Go ≥ 0 or projected onto the orthogonal complement of the OOD regularization gradient Go otherwise, as stated in Eq. (4). The authors evaluate on ImageNet-1K as ID data and four OOD datasets (iNaturalist, SUN, Places, Texture) under one-shot and four-shot settings with CLIP ViT-B/16 and ResNet-50, reporting FPR95, AUROC, ID classification accuracy, and training time. They report that GaCoOp outperforms LoCoOp on FPR95, improves ID accuracy over CoOp, and is substantially faster than ID-like.","tokens_in":10,"tokens_out":7081,"duration_ms":119661,"significance":"The proposed update rule is simple and computationally cheap, and the FPR95 improvements over LoCoOp in Table I are consistent across datasets; the training-time advantage over ID-like (1h48m vs. 23h51m in Table II) is practically meaningful. If the mechanism were properly supported, the paper would be a useful incremental contribution to few-shot OOD prompt tuning. However, the current manuscript does not provide direct evidence for the claimed gradient-conflict-mitigation mechanism, and the update in Eq. (4) does not actually optimize the OOD regularization term as part of a joint objective. The empirical evaluation also lacks error bars and multiple seeds, leaving the central claim unsupported in its current form.","major_comments":[{"comment":"The proposed update does not implement optimization of the objective in Eq. (3). In both branches of Eq. (4), Ggacoop is a function of Gi only; Go is used only as a reference direction for deciding whether to project. Consequently, the OOD regularization gradient never enters the update as an optimization direction. The first-order change of Lood along the update is Go·Ggacoop, which is Gi·Go in the acute branch (nonnegative, so Lood tends to increase) and 0 in the obtuse branch. Thus GaCoOp never descends on Lood. This is inconsistent with the paper's claim that it alleviates the conflict by jointly optimizing ID classification and OOD regularization. The authors should either reframe the method as a constrained modification of CoOp in which the OOD term acts only as a projection reference and compare against simply removing λLood from Eq. (3), or provide evidence that the projection improves OOD detection through a different mechanism.","section":"III-B, Eq. (4)"},{"comment":"The paper reports no ablation or diagnostic that directly supports the gradient-conflict mechanism. There are no statistics on how often the obtuse branch triggers, no distribution of Gi·Go during training, no trajectory of Lood or Lcoop, and no comparison to LoCoOp without the OOD regularization term. Without such evidence, the observed FPR95 improvements could be explained by the projection acting as an ad hoc regularizer of the ID gradient, or by the method partially ignoring a potentially harmful OOD term, rather than by alignment of two co-optimized losses. This missing experiment is load-bearing for the paper's central claim.","section":"IV-C and Table I"},{"comment":"All reported results are single-run point estimates with no error bars, multiple seeds, or significance tests. Prompt tuning is stochastic due to prompt initialization, batch sampling, and data shuffling, and many reported differences (e.g., ID accuracy 69.63 vs. 69.38 in Table II, AUROC differences around one percentage point in Table I) are likely within seed variance. The authors should report mean ± standard deviation over at least 3–5 seeds and, ideally, paired tests for the key comparison against LoCoOp.","section":"IV-A, Tables I and II"},{"comment":"The claim that the orthogonal projection 'can not override the Gi' and that the resulting direction is 'non-conflicting' is not justified as stated. Orthogonality to Go at a single step only ensures zero first-order change of Lood along the update; since both Gi and Go are functions of the current prompt parameters and change after every step, a direction that is orthogonal at step t can become conflicting at step t+1. The paper provides no convergence or safety analysis, and the premise that Go is a trustworthy reference direction is not examined beyond the final metrics.","section":"III-B"}],"minor_comments":[{"comment":"The phrases 'inner ID samples' and 'ImageNed-1k' appear to be typos; please use 'in-distribution ID samples' and 'ImageNet-1K'.","section":"Abstract and Section IV-B"},{"comment":"The symbol K is used in 'K ID-irrelevant region indices' but never defined, and pj should be defined precisely, e.g., as the softmax probability over the j-th region.","section":"III-A, Eq. (2)"},{"comment":"The caption for (b) says 'If Gi is aligned with Go', but the condition in Eq. (4) is Gi·Go ≥ 0, which also includes orthogonal directions; the caption and the surrounding text should be consistent on the angle condition.","section":"Fig. 1"},{"comment":"The table caption does not state the number of shots; the text mentions four-shot results, but the caption should be self-contained, and it should be clear that MCM is a zero-shot method.","section":"Table III"},{"comment":"The repository link is provided, but no code is currently released; please make the code available or state the intended release date for reproducibility.","section":"Abstract"},{"comment":"The projection rule is described as 'inspired by [27]', but the methodological overlap with Prompt-Aligned Gradient is substantial; the authors should state explicitly what is new relative to [27] beyond the application to OOD detection with LoCoOp.","section":"I and III-B"}],"recommendation":"major_revision","confidential_remarks":"The paper presents an interesting empirical direction, but the conceptual gap between Eq. (4) and the claimed two-objective optimization is substantial. For a journal publication, I would require the ablations and seed variance before endorsing it. The contribution may be more modest than claimed once the OOD term is removed; if the authors can show that the projection itself matters beyond simply omitting Lood, the work could be acceptable. No ethical concerns; the main reproducibility issue is the unreleased code."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What should you know first: this paper is a straightforward application of a known gradient-projection update (PCGrad/ProGrad) to LoCoOp's two losses for few-shot OOD detection. The empirical results are decent — FPR95 improves consistently over LoCoOp on the ImageNet benchmark — but the paper's own explanation of what the update does doesn't hold up under inspection.\n\nThe genuinely new part is the task application: taking the projection rule from ProGrad and using the OOD regularization gradient as the reference direction when tuning prompts for LoCoOp. The experiments follow the standard protocol, include a CNN backbone check, and show real gains on FPR95 (e.g., 40.17→29.59 one-shot, 36.95→29.40 four-shot averaged over the four OOD sets). That is a useful data point. The paper also reports ID accuracy improvement, which is a nice touch.\n\nThe soft spots are more than cosmetic. First, as the stress-test note observes, Eq. (4) never uses Go as an optimization direction. In both branches the update is a function of Gi alone; Go only decides whether to project. So the method does not actually optimize the two-term objective in Eq. (3) — it trains on Lcoop with a projection constraint. That means the 'gradient conflict mitigation' framing is inaccurate; the mechanism is closer to a regularizer on the ID gradient. Second, the update rule is exactly ProGrad's, which the authors cite [27] but do not position as the source of the algorithm. The novelty is in the application, not the rule. Third, there are no error bars or multiple seeds, and the AUROC numbers are mixed against ID-like, so the headline 'great performance' is only partially supported. The paper would benefit from reporting the frequency of the obtuse branch, angle distributions, and an ablation that actually adds the λGo term back, so we can see whether the projection or the omission of Lood drives the gains. The claim that the projection 'discards conflicting knowledge' assumes Go is a trustworthy reference; that premise is asserted, not demonstrated.\n\nWho this is for: researchers working on few-shot OOD detection with CLIP who want a simple, cheap improvement over LoCoOp. As a conference submission it is borderline: the empirical study is solid enough to warrant a serious referee, but the paper needs a rewrite of the mechanism section and at least a reproducibility check before acceptance. I'd send it to review, with a clear request for seeds, code, and the missing ablations.","headline":"Straightforward application of a known gradient-projection rule (ProGrad/PCGrad) to LoCoOp, with decent FPR95 gains but a mechanism story that doesn't survive close reading.","tokens_in":8839,"tokens_out":3516,"would_cite":false,"duration_ms":32709,"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":"The paper claims that projecting the in-distribution gradient away from the OOD-regularization gradient removes the conflict in few-shot prompt tuning.","keywords":["few-shot out-of-distribution detection","prompt learning","gradient projection","gradient conflict","CLIP","context optimization","ImageNet OOD benchmark","OOD regularization"],"falsifier":"Run GaCoOp with the projection branch always disabled and with it always enabled, and record the fraction of training steps where $\\mathbf{G}_i \\cdot \\mathbf{G}_o < 0$ actually occurs. If removing the projection leaves the final FPR95 and ID accuracy unchanged, the conflict it targets was not the limiting factor; if the conflicting fraction is near zero, the projection is not the source of the gain.","tokens_in":7883,"feed_emoji":"🎯","tokens_out":9239,"duration_ms":74399,"temperature":0.7,"pith_summary":"Few-shot out-of-distribution detection has to solve two tasks at once: keep classifying in-distribution images correctly while using the background regions of the few labeled training images as fake OOD data and pushing their features away from the ID text embeddings. LoCoOp, the baseline this paper starts from, treats this as one loss $L = L_{\\text{coop}} + \\lambda L_{\\text{ood}}$, and the paper claims that these two terms often pull the learnable prompt in opposite directions when the pretrained vision-language model is biased about which image regions are background. GaCoOp fixes this: at each update step it computes the ID classification gradient $\\mathbf{G}_i$ and the OOD regularization gradient $\\mathbf{G}_o$, and if the angle between them is obtuse it discards the component of $\\mathbf{G}_i$ parallel to $\\mathbf{G}_o$, keeping only the orthogonal part. The authors claim this restores cooperation between the objectives, and their ImageNet experiments show lower FPR95 and higher ID accuracy than the baseline at comparable training cost. If true, the result matters because it turns a known source of instability in few-shot OOD detection into a simple one-line update rule.","feed_headline":"Project away the conflict to sharpen few-shot OOD detection","feed_subtitle":"Keeping only the in-distribution gradient that agrees with OOD regularization lowers FPR95 and raises ID accuracy.","key_machinery":"The load-bearing mechanism is the gradient projection rule of Eq. (4), applied to the context vectors of the prompt. The rule treats the OOD regularization gradient $\\mathbf{G}_o$ as a fixed reference direction and modifies only the ID gradient: an acute dot product means keep $\\mathbf{G}_i$ unchanged, while an obtuse dot product means replace $\\mathbf{G}_i$ by its projection onto the subspace orthogonal to $\\mathbf{G}_o$. This projection is what lets the two losses cooperate without reweighting them.","core_discovery":"The central claim is embodied in Eq. (4): the update direction is $\\mathbf{G}_i$ when $\\mathbf{G}_i \\cdot \\mathbf{G}_o \\ge 0$, and otherwise $\\mathbf{G}_i - (\\mathbf{G}_i \\cdot \\mathbf{G}_o / \\|\\mathbf{G}_o\\|^2)\\,\\mathbf{G}_o$. The paper's way of stating it is that the in-distribution gradient decomposes into a non-conflicting orthogonal part and a potentially conflicting parallel part; when the parallel part points against $\\mathbf{G}_o$, dropping it prevents OOD regularization from corrupting ID classification. The authors claim this gradient aligned context optimization improves OOD detection on the ImageNet benchmarks and improves ID classification accuracy on ImageNet over CoOp.","pith_inferences":["The paper does not report how often the conflict condition $\\mathbf{G}_i \\cdot \\mathbf{G}_o < 0$ actually fires; if that fraction is low, the reported gains may come from the unchanged $\\mathbf{G}_i$ updates rather than from the projection, a distinction a follow-up could test.","The angle-only projection ignores gradient magnitudes and curvature, so the same rule may need a step-size correction or a margin on the dot product when applied to larger prompts or to other two-loss objectives.","The paper's closing belief that an unbiased classifier is also a good OOD detector implies a calibration claim: GaCoOp should also reduce in-distribution overconfidence, which could be checked with expected calibration error on ImageNet.","The same decomposition could apply to any prompt-tuning setup that combines a task loss with a regularizer, for example supervised contrastive learning plus entropy maximization."],"forward_implications":["On the ImageNet OOD benchmark, the one-shot average FPR95 drops from 40.17 for LoCoOp to 29.59 for GaCoOp, and the four-shot average drops from 36.95 to 29.40.","ID classification accuracy on ImageNet-1K reaches 69.63 percent, slightly above CoOp's 69.38 percent, while training takes only about 30 minutes longer than CoOp.","The same projection rule transfers to a ResNet-50 backbone, improving average FPR95 from 45.62 for LoCoOp to 33.74 and average AUROC to 91.62.","Because the projection only changes the update direction on conflicting steps, the rule can be grafted onto any prompt-tuning objective of the form $L_{\\text{coop}} + \\lambda L_{\\text{ood}}$."],"supporting_citations":[{"why":"Supplies CLIP, the pretrained vision-language model whose image-text similarity defines both classification and OOD regularization.","marker":"[3]"},{"why":"Defines the baseline LoCoOp and the OOD regularization loss $L_{\\text{ood}}$ whose gradient is $\\mathbf{G}_o$.","marker":"[9]"},{"why":"Defines CoOp and the cross-entropy classification loss $L_{\\text{coop}}$ whose gradient is $\\mathbf{G}_i$.","marker":"[18]"},{"why":"Provides the gradient projection idea that GaCoOp adapts for aligning prompt gradients.","marker":"[27]"},{"why":"Supplies the strongest few-shot OOD detection comparison and the ImageNet evaluation protocol.","marker":"[10]"},{"why":"Supplies the zero-shot OOD detection baseline MCM and the AUROC/FPR95 evaluation metrics.","marker":"[4]"},{"why":"Supplies ImageNet-1K as the in-distribution dataset for all reported experiments.","marker":"[28]"}],"fun_headline_variants":["GaCoOp projects away conflicting gradients for better OOD detection","Drop conflicting gradients to sharpen few-shot OOD detection","Gradient alignment reduces conflict in few-shot OOD detection","Project gradients to cut conflict and boost OOD detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the angle between the two gradient vectors at a single step is a complete and trustworthy measure of conflict, so removing the component of the in-distribution gradient that points against the OOD gradient always helps and never hurts.","fun_headline_variants_meta":{"raw":{"variants":["GaCoOp projects away conflicting gradients for better OOD detection","Drop conflicting gradients to sharpen few-shot OOD detection","Gradient alignment reduces conflict in few-shot OOD detection","Project gradients to cut conflict and boost OOD detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000187,"raw_usage":{"total_tokens":1291,"prompt_tokens":869,"completion_tokens":422,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":485,"completion_tokens_details":{"reasoning_tokens":356}},"tokens_in":485,"tokens_out":422,"duration_ms":4409,"temperature":1.0,"reasoning_tokens":356,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:57:15.894341+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GaCoOp with the projection branch always disabled and with it always enabled, and record the fraction of training steps where $\\mathbf{G}_i \\cdot \\mathbf{G}_o < 0$ actually occurs. If removing the projection leaves the final FPR95 and ID accuracy unchanged, the conflict it targets was not the limiting factor; if the conflicting fraction is near zero, the projection is not the source of the gain.","supporting_citations":[{"cited_title":"Locoop: Few-shot out- of-distribution detection via prompt learning,","cited_arxiv_id":null,"evidence_quote":"Defines the baseline LoCoOp and the OOD regularization loss $L_{\\text{ood}}$ whose gradient is $\\mathbf{G}_o$."},{"cited_title":"Id-like prompt learning for few-shot out-of-distribution detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the strongest few-shot OOD detection comparison and the ImageNet evaluation protocol."}],"review_version":1}