{"id":"56d6a36f-ef4a-4986-adad-624f6cc9ec46","arxiv_id":"2507.11969","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GS-Bias improves zero-shot CLIP accuracy by learning per-image global and spatial biases at the output-score level, achieving reported state-of-the-art test-time adaptation with far lower memory use than TPT.","lead":"A new test-time adaptation method for vision-language models adds two small trainable bias vectors directly to the model output scores, one learned from augmented views of the image and one from selected image regions. On 15 benchmarks it reports higher accuracy than prior test-time adaptation methods while using about 6.5% of the memory of a leading prompt-tuning approach.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported SOTA gains are confounded by prompt ensembling applied only to GS-Bias and by hyperparameters (K, α, β) tuned on the evaluation benchmarks; a validation-locked, ensemble-matched replication is needed before the central accuracy claim can be trusted.","rationale":"The reader identified the spatial-region selection as the weakest assumption. I partially agree, but find a more immediate load-bearing problem: the empirical SOTA claim is not separable from evaluation-protocol choices. The paper's own Table 1 shows no-ensemble GS-Bias is nearly tied with TPT; the advertised +2.23/+2.72 margins come from adding prompt ensembling to only GS-Bias. Similarly, hyperparameters are chosen using the test benchmarks, including the just-so K=16 justification in Table 7. These are fixable with validation-based selection and ensemble-matched baselines, so the right outcome is still CONDITIONAL rather than acceptance or rejection. The spatial-region concern remains relevant: if the category-aware top-K selection is what makes the spatial bias work, a random-patch ablation would show it; that should be part of the same validation-locked replication.","tokens_in":20610,"tokens_out":10031,"duration_ms":122749,"concrete_test":"Run a preregistered evaluation with: (a) all baselines (TPT, DiffTPT, MTA) given the same prompt-ensemble option as GS-Bias; (b) all hyperparameters (K, α, β, BS, ρ, steps) selected on a held-out validation split per benchmark (e.g., ImageNet val for the DG set and a 20% split of each cross-dataset set), never on the test split; (c) at least 3 seeds or bootstrap confidence intervals. If GS-Bias+E remains ahead of the ensemble-matched baselines by more than the interval width on both benchmarks, the central claim stands; otherwise the reported margins are evaluation artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: GS-Bias is state-of-the-art. To hold, the evaluation must not select hyperparameters on the test labels and must compare like-for-like. Both fail. In §4.2, the headline gains (67.03 vs 64.80 on cross-dataset; 63.01 vs 60.29 OOD) are reported for 'GS-Bias + E.', where E. is CLIP's prompt ensemble, while the TPT/DiffTPT rows do not use an ensemble; the unensembled GS-Bias is only 64.86 (vs 64.80 TPT) on cross-dataset and 60.69 OOD (vs 60.29 TPT), and is below MTA+E (62.53 OOD). Second, K=16 is justified in Table 7 by counting 'significant regions' on the same 11 datasets used for evaluation, and Fig. 3 selects K on those datasets; α and β are chosen from ablations on the evaluation benchmarks (Table 5, Fig. 6). No held-out validation split or error bars are reported, so the small remaining gains after removing the ensemble could be selection noise. The released code makes the issue testable, but as reported the advertised margins are not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GS-Bias, a test-time adaptation method for vision-language models that learns two additive bias vectors: a global bias shared across augmented views and a spatial bias derived from top-K regions in the vision encoder's spatial feature map. Both biases are optimized by entropy minimization at the model output, intentionally avoiding backpropagation through the VLM. The paper reports state-of-the-art results across 10 cross-dataset benchmarks and 4 ImageNet variants, with substantially lower memory usage than TPT, and releases the code.","tokens_in":20907,"tokens_out":5934,"duration_ms":62904,"significance":"If the accuracy claims hold, GS-Bias offers an attractive efficiency/accuracy trade-off for TTA, and the output-level optimization is a conceptually clean alternative to full prompt tuning. The released code and the efficiency analysis are concrete strengths. However, the reported margins over TPT are partly due to prompt ensembling applied only to GS-Bias and to hyperparameters selected on the evaluation benchmarks; the unensembled gains are small. These issues currently weaken the central state-of-the-art claim, although the method's core idea appears sound and the concerns are addressable in revision.","major_comments":[{"comment":"The text repeatedly states that biases are added to logits, but pCLIP in Eq. (1) and p(y_c | F^i_s) in Eq. (9) are softmax probabilities. Adding Bg or Bs to these probability vectors yields distributions that do not sum to one, and the entropy objectives in Eqs. (3) and (13) are then not defined over a valid probability distribution. Please clarify whether the biases are applied before or after the softmax, revise the equations and terminology accordingly, and explain how the optimization remains well-posed in the actual implementation.","section":"Section 3.2, Eqs. (6), (8), (12), (15)"},{"comment":"The headline improvements over TPT (2.23% cross-dataset, 2.72% OOD) compare GS-Bias + E. with a TPT baseline that does not use prompt ensembling. The unensembled GS-Bias achieves only 64.86 vs 64.80 in Table 1 and 60.69 vs 60.29 in Table 2, which is a much smaller difference and within the range of possible random variation. To support the SOTA claim, the primary comparison should be ensemble-matched (for example, unensembled vs unensembled, or both with the same ensemble), and the remaining margin should be demonstrated to be robust.","section":"Section 4.2, Tables 1 and 2, abstract"},{"comment":"Hyperparameters K, alpha, beta, and the number of TTA steps are selected by ablations on the same 11 datasets used for evaluation (Fig. 3, Table 5, Fig. 4, Fig. 6, Table 7). Tuning on the test benchmarks can inflate reported accuracy. The authors should adopt a held-out validation procedure, for example tuning on a subset of ImageNet or on a separate dataset, and then fixing the hyperparameters for the reported results. The Limitation statement in A.4 acknowledges reliance on empirical hyperparameters, but the main experimental claims still rest on these test-set-tuned choices.","section":"Section 4.3, Appendix A.3, Appendix A.4"},{"comment":"No error bars, standard deviations, or multiple seeds are reported. Given that the unensembled cross-dataset margin over TPT is only 0.06 percentage points (64.86 vs 64.80), single-run results are insufficient to establish the improvement. Report mean and standard deviation over at least three runs, or provide a significance test, especially for the average accuracies.","section":"Section 4.2, Tables 1 and 2"}],"minor_comments":[{"comment":"The phrase \"logits outputed\" should be \"logits output\"; please also check the spelling and grammar throughout the manuscript.","section":"Abstract and Section 3.2"},{"comment":"\"As illustrated in Fig. 1 (c)\" should reference Fig. 2 (c), which contains the GS-Bias diagram.","section":"Section 3.2, third paragraph"},{"comment":"The text \"as per Eq. 2\" appears to be a wrong cross-reference; the self-entropy is defined in Eq. (3). Please correct the reference.","section":"Section 3.2, after Eq. (6)"},{"comment":"The notation \"Top-K(M^i)\" should be specified precisely, e.g., as the indices of the K largest entries of the vector M.","section":"Eq. (11)"},{"comment":"The phrase \"filters high-uncertainty logits smaller than θ\" is confusing; it should say that views with entropy above θ are filtered out, matching the indicator function in Eq. (2).","section":"Section 3.2, Eq. (6) description"},{"comment":"The notation \\tilde{K}_a and the normalization used for the significance threshold (0.1) should be defined more clearly in the text accompanying the table.","section":"Appendix A.3, Table 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is likely to be of interest to the test-time adaptation community, and the code release is a positive feature. However, the validation protocol—test-set hyperparameter tuning and unmatched ensembling—needs to be corrected before the state-of-the-art claim is credible. If the authors can provide validation-locked hyperparameters, ensemble-matched baselines, and variance estimates, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the GS-Bias paper. The core idea is genuinely neat: instead of tuning prompts or visual features, add two learnable bias vectors to the output logits and optimize them with entropy minimization over augmented views and selected spatial patches. That gets most of TPT's accuracy at a fraction of the memory cost (about 6.5% on ImageNet). The spatial-bias part, using top-K patch-class relevance to pick regions, is a sensible extension of prompt-tuning insights to training-free TTA. The paper is clearly written, ships code, and evaluates across 15 datasets including a ResNet50 variant. That is real work.\n\nThe headline numbers deserve scrutiny. The 2.23% and 2.72% improvements over TPT come from GS-Bias plus prompt ensembling, while TPT is reported without ensembling (which the paper justifies). Unensembled GS-Bias is only +0.06 on cross-dataset and +0.40 on OOD, and below MTA+E on OOD. However, GS-Bias+E does beat MTA+E by about 1.1 and 0.5, so the SOTA claim is not baseless, but the abstract should say \"with ensembling.\" More worrying is hyperparameter selection: K, alpha, and beta are tuned on the same test benchmarks via ablations (Figure 3, Table 5, Table 7). The limitation appendix admits the empirical reliance, but with margins this small, selection noise could explain the remaining advantage. Also, the equations add the biases to softmax probabilities while the text says logits; that is a notation or implementation mismatch that needs fixing. If it is logits, say so; if it is probabilities, renormalization is needed. No error bars or seeds are reported, which is common but worth noting.\n\nOverall, the method is a legitimate contribution and the efficiency story is solid. The accuracy story is plausible but not nailed down as reported. This is the kind of paper a serious referee should see; the issues are addressable, not fatal. I would accept it for peer review, with the main asks being: report all baselines with and without ensembling, choose hyperparameters on a held-out validation split, and reconcile the logit/probability notation. If those check out, the method earns its place.","headline":"A useful logit-level TTA idea with real efficiency gains, but the headline accuracy margins rest on an ensemble apples-to-oranges comparison and hyperparameters tuned on the test benchmarks.","tokens_in":21455,"tokens_out":3064,"would_cite":true,"duration_ms":33979,"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":"GS-Bias adapts CLIP at test time by adding two learnable bias vectors to the output logits, minimizing prediction entropy across augmented views and class-relevant regions, and beats prompt tuning on 15 benchmarks at 6.5% of its memory.","keywords":["test-time adaptation","vision-language models","CLIP","logit-level bias","entropy minimization","spatial region selection","cross-dataset generalization","domain generalization"],"falsifier":"A direct test is to take a dataset where the object occupies an atypical share of the image (very small foreground or a large object nearly filling the frame) and compare GS-Bias against its global-bias-only variant: if the top-$K$ region selection falls mostly on background or splits the object, the spatial bias should inject a class offset that lowers accuracy below the global-only version. The paper's own Table 7 already shows the number of genuinely significant regions varies widely across datasets, about 12 for EuroSAT versus 19 for ImageNet, so the fixed $K=16$ setting is the natural place to look for the breakdown.","tokens_in":20398,"feed_emoji":"⚡","tokens_out":20408,"duration_ms":183293,"temperature":0.7,"pith_summary":"The paper claims that test-time adaptation of CLIP can be carried out entirely at the output layer: instead of tuning text prompts or refining visual features, GS-Bias learns two small additive bias vectors, a global one and a spatial one, and adds them directly to the frozen model's logits. The global bias is trained so that several augmented views of a single test image predict consistently, and the spatial bias is trained so that the image's most class-relevant regions predict consistently. Since gradients flow only into these two vectors, adaptation uses about 6.5% of the memory and runs roughly ten times faster than the leading prompt-tuning method, while reporting higher average accuracy on cross-dataset and out-of-distribution benchmarks. If true, this shows the consistency signal that drives test-time adaptation does not require modifying the model's internals at all.","feed_headline":"Two learnable biases beat prompt tuning at 6.5% of its memory","feed_subtitle":"Adding a global and a spatial bias to the frozen logits lifts accuracy on 15 benchmarks at a fraction of the cost.","key_machinery":"The load-bearing object is a pair of learnable bias vectors, $B_g \\in \\mathbb{R}^{1\\times C}$ and $B_s \\in \\mathbb{R}^{1\\times C}$, added at the logits stage to the frozen CLIP output distribution (Eq. 15). $B_g$ is fit by entropy minimization over confidence-filtered augmented views of the test image (Eqs. 6-7), encoding whole-image semantic consistency; $B_s$ is fit by entropy minimization over the average scores of the top-$K$ spatial regions selected by a category-aware relevance map $M = \\frac{1}{C}\\sum_c \\mathrm{Softmax}(F_s (F_t^c)^T)$ (Eqs. 10-14), encoding regional semantic coherence. Because the only parameters touched by backpropagation are these two vectors, the mechanism's cost is a few matrix-vector operations at the output layer rather than a full forward-backward pass through the vision-language model, which is what yields the reported ten-fold speedup and 6.5% memory footprint relative to prompt tuning.","core_discovery":"GS-Bias's central claim is that when CLIP encounters a single test image, its per-image accuracy can be improved by two learnable, zero-initialized bias vectors added at the prediction level, $p_{\\text{GS-Bias}}(y|x) = p_{\\text{CLIP}}(y|x) + B_g + B_s$ (Eq. 15). The global bias $B_g$ is shared across augmented views of the image and updated by minimizing the entropy of the confidence-filtered average of the view logits (Eqs. 6-7); the spatial bias $B_s$ is shared across the top-$K$ spatial regions most relevant to the target classes, selected by the category-aware map $M = \\frac{1}{C}\\sum_c \\mathrm{Softmax}(F_s (F_t^c)^T)$, and updated by minimizing the entropy of their average class scores (Eqs. 10-14). Both biases are optimized with a handful of lightweight steps, and the final prediction adds them to the untouched CLIP logits, so the pretrained model itself is never modified. The paper reports this recipe reaches 67.03% average accuracy on ten cross-dataset benchmarks (vs. 64.80% for TPT) and 63.01% OOD average on four ImageNet variants (vs. 60.29% for TPT), while using about 6.5% of TPT's memory on ImageNet.","pith_inferences":["Because the bias vectors scale with the number of classes, an extension the paper leaves implicit is factorizing or sharing structure across $B_g$ and $B_s$ to keep the overhead flat on very large label sets such as ImageNet-21K.","The mechanism is not obviously CLIP-specific: any frozen softmax classifier exposes the same output layer, so the global-plus-spatial bias recipe could be tested on other multimodal encoders or unimodal classifiers with patch-like features.","A testable follow-up would recompute the category-aware region map after the global bias has been updated, letting region selection and bias learning co-adapt within one image instead of fixing the selected regions once."],"forward_implications":["With only the two bias vectors updated, adaptation on ImageNet runs at roughly 12 FPS with about 1.3 GB of memory, compared with about 1.4 FPS and 20 GB for TPT, making per-image adaptation practical in streaming settings.","Because the pretrained logits are preserved and the biases start at zero, GS-Bias with ensemble prompts improves on the ensemble-prompt CLIP baseline on all ten cross-dataset datasets, a consistency none of the compared optimizers achieves.","The ablation attributes the gains to both terms: the global bias alone lifts the cross-dataset average to 65.77, the spatial bias alone to 66.07, and both together to 67.03, so whole-image and region-level consistency capture different failure modes.","The recipe transfers to other visual backbones: with ResNet50 features as the spatial representation, GS-Bias reports 45.93% OOD average on domain generalization, the best among the methods compared in that table."],"supporting_citations":[{"why":"Supplies the test-time prompt-tuning baseline GS-Bias must beat and the entropy-plus-confidence-filtering objective over augmented views that the global bias reuses at the logit level.","marker":"(Shu et al., 2022)"},{"why":"Provides the training-free visual-optimizer baseline (MTA) whose speed GS-Bias matches while reporting higher accuracy.","marker":"(Zanella & Ben Ayed, 2024)"},{"why":"Supplies CLIP itself, the frozen model whose output logits and spatial patch features the two biases are added to and selected over.","marker":"(Radford et al., 2021a)"},{"why":"DiffTPT, the diffusion-augmented prompt-tuning baseline that GS-Bias exceeds without any generative augmentation.","marker":"(Feng et al., 2023)"},{"why":"LoCoOp, cited as evidence that enriching spatial features carries class-relevant information, motivating the spatial bias.","marker":"(Miyai et al., 2024)"},{"why":"GalLOP, the global-and-local prompt study used to justify learning both whole-image and regional consistency.","marker":"(Lafon et al., 2024)"},{"why":"Defines the ViT backbone whose patch-token features supply the spatial representation used for region selection.","marker":"(Dosovitskiy, 2020)"}],"fun_headline_variants":["Two biases beat prompt tuning at 6.5% memory cost","Add two biases to CLIP logits for cheap test-time boost","Zero-shot CLIP gets 2.7% gain with 6.5% memory","Global and spatial biases: tiny tweaks, big test-time gains","Two biases top 15 benchmarks at 6.5% memory cost"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"GS-Bias depends on the assumption that the top-$K$ spatial regions flagged as most relevant to the class labels actually carry the class-discriminative information, so that making their average prediction more confident also improves the whole-image prediction; in the paper, $K=16$ is justified after the fact by counting 'significant' regions on the same datasets being evaluated.","fun_headline_variants_meta":{"raw":{"variants":["Two biases beat prompt tuning at 6.5% memory cost","Add two biases to CLIP logits for cheap test-time boost","Zero-shot CLIP gets 2.7% gain with 6.5% memory","Global and spatial biases: tiny tweaks, big test-time gains","Two biases top 15 benchmarks at 6.5% memory cost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000833,"raw_usage":{"total_tokens":3721,"prompt_tokens":1113,"completion_tokens":2608,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":729,"completion_tokens_details":{"reasoning_tokens":2511}},"tokens_in":729,"tokens_out":2608,"duration_ms":18780,"temperature":1.0,"reasoning_tokens":2511,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:56:55.231823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test is to take a dataset where the object occupies an atypical share of the image (very small foreground or a large object nearly filling the frame) and compare GS-Bias against its global-bias-only variant: if the top-$K$ region selection falls mostly on background or splits the object, the spatial bias should inject a class offset that lowers accuracy below the global-only version. The paper's own Table 7 already shows the number of genuinely significant regions varies widely across datasets, about 12 for EuroSAT versus 19 for ImageNet, so the fixed $K=16$ setting is the natural place to look for the breakdown.","supporting_citations":[{"cited_title":"Test-time prompt tuning for zero-shot generalization in vision-language models","cited_arxiv_id":null,"evidence_quote":"Supplies the test-time prompt-tuning baseline GS-Bias must beat and the entropy-plus-confidence-filtering objective over augmented views that the global bias reuses at the logit level."},{"cited_title":"and Ben Ayed, I","cited_arxiv_id":null,"evidence_quote":"Provides the training-free visual-optimizer baseline (MTA) whose speed GS-Bias matches while reporting higher accuracy."},{"cited_title":"Diverse data augmentation with diffusions for effective test-time prompt tuning","cited_arxiv_id":null,"evidence_quote":"DiffTPT, the diffusion-augmented prompt-tuning baseline that GS-Bias exceeds without any generative augmentation."},{"cited_title":"Locoop: Few-shot out-of-distribution detection via prompt learning","cited_arxiv_id":null,"evidence_quote":"LoCoOp, cited as evidence that enriching spatial features carries class-relevant information, motivating the spatial bias."},{"cited_title":"Gallop: Learning global and local prompts for vision-language models","cited_arxiv_id":null,"evidence_quote":"GalLOP, the global-and-local prompt study used to justify learning both whole-image and regional consistency."}],"review_version":1}