{"id":"1f446706-c451-43ec-b380-41d56b8afeb9","arxiv_id":"2506.09399","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A dynamic covariance adjustment that contracts the Mahalanobis metric along each test feature's residual-space direction improves OOD detection over static priors on several backbones.","lead":"This paper proposes a test-time tweak to OOD detection: it shrinks the covariance matrix along the direction of each new image's features before measuring Mahalanobis distance. The tweak improves separation between known and unknown images across CIFAR and ImageNet models, including self-supervised DINO.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The score in Eq. 3 is not guaranteed well-defined: the paper's own p>1 regime makes M(f) indefinite, and Theorem 4.2 only certifies one anchor vector, not every class residual used by the min.","rationale":"The reader correctly identifies Theorem 4.2's conditions as under-verified, but the deeper issue is that the theorem, even if true, does not support M(f) as a metric. In the p>1 regime the paper explicitly embraces (Section 4.3, Figure 6), M(f) has a negative eigenvalue, so it cannot define a distance. The nonnegativity guarantee in Theorem 4.2 covers a single quadratic form (f-a)^T M(f)(f-a); Eq. 3 uses the same indefinite M(f) for all class differences r_i, and the min requires every sqrt argument to be real. The appendix validates only the closest class mean, leaving the other N_c-1 terms uncertified. This does not refute the empirical gains, but it makes the theoretical framing unreliable and keeps the central claim conditional. A secondary inconsistency is Section 5.3's claim of state-of-the-art on all four ImageNet models versus Table 2, where WDiscOOD beats the method on ViT (AUROC 94.41 vs 94.27; FPR95 26.35 vs 26.94) and VIM beats it on Swin-B FPR95 (51.33 vs 51.89), and the average row excludes WDiscOOD on Swin-B. Together these issues support caution but not rejection; the reader's CONDITIONAL verdict remains appropriate.","tokens_in":24459,"tokens_out":12985,"duration_ms":132009,"concrete_test":"Instrument the released code to log, for every ID and OOD feature in each benchmark, the N_c values r_i^T M(f) r_i before the sqrt in Eq. 3, and the minimum eigenvalue of M(f). If any value is negative while reported scores are real, the published implementation deviates from Eq. 3; if all values are nonnegative, verify whether the p>1 branch occurs and whether the theorem's condition holds for all classes, not just the closest one. Additionally, recompute Table 2's ViT row to confirm the reported AUROC/FPR95, since Section 5.3's SOTA claim is contradicted by the table itself.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central score (Eq. 3) is s(f) = -min_i sqrt(r_i^T M(f) r_i), with M(f) = (Σ_R - B^T a a^T B)^{-1}. Theorem 4.2 analyzes d(f) = (f-a)^T (Σ - f f^T)^{-1}(f-a) for a single anchor a, and its p>1 branch requires (s-1)^2 ≤ (p-1)(q-1). For p>1, however, Σ - f f^T has a negative eigenvalue (Sherman-Morrison determinant det(Σ)(1-p) < 0), so M(f) is indefinite and does not induce a metric. Section 4.3 explicitly embraces p≫1 and q≫1, and Figure 6 confirms p values far above 1. The theorem proves nonnegativity only for the particular vector f-a, not for r_i = f-μ_i for every class i. The min in Eq. 3 requires every r_i^T M(f) r_i to be nonnegative; if any is negative, the score is undefined (and sqrt would produce NaN in Algorithm 1). Appendix B.1 checks s≥0 only for the closest class mean (Figure 8), not for all N_c class residuals. Thus well-definedness of the reported score rests on an unverified, class-wise condition, and the paper's stated theoretical guarantee does not cover the actual computation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dynamic Covariance Calibration (DCC), a post-hoc out-of-distribution detection score that dynamically adjusts the within-class covariance matrix in the residual space of the training feature distribution using the real-time input feature. The score is s(f) = -min_i sqrt(r_i^T (Sigma_R - B^T a a^T B)^{-1} r_i), with r_i = f - mu_i, a the residual-space coefficient vector of f, B the residual basis, and Sigma_R the within-class covariance. The authors justify the construction with a matrix-perturbation theorem (Theorem 4.2) and evaluate on CIFAR (DenseNet, WideResNet) and ImageNet-1k (ViT, ResNet-50, Swin-B, DeiT, DINO) benchmarks, reporting strong AUROC and FPR95 numbers and claiming state-of-the-art performance.","tokens_in":24770,"tokens_out":4832,"duration_ms":49096,"significance":"The core idea is simple, test-time only, and computationally cheap, and the empirical study is broad, covering multiple architectures including a self-supervised DINO model, with code released. These are genuine strengths. However, the theoretical guarantee as stated does not cover the actual algorithm, because Theorem 4.2 certifies only a single anchor vector while the score takes a minimum over all class residuals; for p>1 the working matrix is indefinite and the score may be undefined. The reported SOTA claim is also overstated for the ViT row in Table 2, and the residual dimension is selected from test performance in Section 5.6. If the well-definedness issue is resolved and the benchmarking is made more rigorous (error bars, fixed validation protocol), the method would be a useful contribution to post-hoc OOD detection.","major_comments":[{"comment":"Theorem 4.2 only proves d(f) >= 0 for a single anchor a, but the score in Eq. (3) requires r_i^T M(f) r_i >= 0 for every class residual r_i = f - mu_i before taking the minimum. For p > 1, the Sherman-Morrison determinant shows Sigma_R - B^T a a^T B is indefinite, so some quadratic forms may be negative and the square root in Algorithm 1 may be undefined. Appendix B.1 (Figure 8) verifies s >= 0 only for the closest class mean, not for all N_c classes, and no figure verifies the actual inequality (s-1)^2 <= (p-1)(q-1). The paper must either prove the condition for every class residual or modify the algorithm (e.g., clipping or explicit positive-definiteness check) so the reported score is always well-defined.","section":"§4.1, §4.3, Algorithm 1"},{"comment":"The residual space dimension k is selected from the test AUROC: Figures 4 and 5 show performance varying with dimensionality, and the main tables report numbers at the best-performing dimension. No error bars, confidence intervals, or multiple-seed runs are reported in the main tables. This makes the comparison with fixed baselines optimistic and prevents assessing whether the reported improvements are statistically significant. Please fix the dimension by a proper validation protocol or report sensitivity and variance.","section":"§5.6, Tables 1-2"},{"comment":"The claim that the method 'achieves state-of-the-art performance on both AUROC and FPR95 across all four pre-trained models' is contradicted by Table 2 itself: on ViT, WDiscOOD has AUROC 94.41 vs ours 94.27 and FPR95 26.35 vs ours 26.94. In addition, the average row excludes WDiscOOD for Swin-B because of NaN scores, so the comparison is not on identical model sets. The SOTA claim should be qualified accordingly.","section":"§5.3, Table 2"},{"comment":"The validity condition of Theorem 4.2 is only empirically motivated, not verified for the score actually computed. Section 4.3 asserts p >> 1 and q >> 1 and that s >= 0, and Section 5.7 shows distributions of p and q, but the condition of the theorem is the joint inequality (s-1)^2 <= (p-1)(q-1); the s values in Appendix B.1 are shown only for the closest class and are not combined with p and q to check the inequality. Since the positive-definiteness of M(f) is load-bearing for the method, this gap should be addressed directly, either by a sharper proof for the min over classes or by a per-class empirical check reported in the main text.","section":"§4.3, Appendix B.1, §5.7"}],"minor_comments":[{"comment":"The heading 'Theorical Proof' should read 'Theoretical Proof'.","section":"Appendix A"},{"comment":"The column header 'AVG' should be 'Avg.', and 'Comparision' is misspelled in the captions of Tables 1 and 8.","section":"Tables 8-9"},{"comment":"The text alternates between 'principle space' and 'principal space'; the standard term is 'principal space'.","section":"§4.2"},{"comment":"The sentence 'the method without considering the real-time features can not perform better than the full method in all scenarios' is ambiguous and should be rephrased; 'can not' should be 'cannot' and the intended meaning clarified.","section":"§5.5"},{"comment":"The statement doubting the validity of WiscOOD's CLIP experiments, based on an unverified possible overlap between the image-text training set and OOD datasets, is speculative and should either be supported with evidence or removed.","section":"Appendix F"},{"comment":"The caption 'The s values on ImageNet-1k pre-trained ResNet-50 and ViT' describes distributions rather than values; also, the x-axis in panel (b) starts at 0, which makes it difficult to confirm that all values are strictly positive.","section":"Figure 8, Appendix B.1"}],"recommendation":"major_revision","confidential_remarks":"The reader's circularity concern (score fitted to test labels) does not hold: no constants are fitted on test data except the residual dimension, which is a legitimate methodological concern but not circularity. The self-citation of Pan et al. 2023, which includes co-author Tan Pan, is a minor conflict-of-interest signal in the model-distribution discussion, but it does not affect the central claim. The main fixable issues are the well-definedness gap between Theorem 4.2 and Eq. (3), the test-based selection of the residual dimension, and the overstated SOTA claim in Table 2."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a plausible new twist on distance-based OOD detection—using the test feature's residual-space projection to shrink the covariance before computing the score—and the experiments are broad enough to take seriously. But the paper overstates SOTA, and the theoretical guarantee does not actually cover the score as computed.\n\nWhat is new: the dynamic covariance adjustment is not in Mahalanobis, VIM, Neco, WDiscOOD, or FDBD. Those methods fit a fixed geometry from training data; here the covariance is modified per test point, restricted to the residual space. The code is released. The evaluation spans DenseNet, WideResNet, ViT, ResNet-50, Swin, DeiT, and DINO on standard CIFAR and ImageNet OOD benchmarks, which is broad. The near-OOD appendix results are a plus, and the ablation is informative.\n\nSoft spots, in proportion: The biggest one is the gap between Theorem 4.2 and the actual score. The theorem certifies nonnegativity of (f-a)^T(Σ - f f^T)^{-1}(f-a) for one anchor a under a condition. But the score takes a min over N_c class means. For p>1, M(f) is indefinite, so some r_i^T M(f) r_i could be negative for other classes; the appendix only verifies the s condition for the closest class mean, not all classes. The paper has not shown the score is well-defined in general. This is a real gap, though empirically the method appears to run without NaNs. Second, Section 5.3 says SOTA on all four ImageNet models, but their own Table 2 shows WDiscOOD beats them on ViT AUROC and FPR95, and VIM/Neco beat them on FPR95 for Swin and DeiT. The average is good; the per-model claim is not. Third, the residual dimension k is chosen with an eye on test-set curves and the main tables have no error bars. These are fixable.\n\nBottom line: this is a legitimate empirical proposal with a clear new mechanism. I would send it to review, but the authors need to fix the class-wise well-definedness argument, correct the SOTA claim, and report how k is selected and how stable the numbers are. I'd cite it as related work on dynamic test-time geometry.","headline":"A genuinely new test-time covariance adjustment for OOD scores with broad experiments, but the well-definedness proof does not cover the min over classes and the SOTA claim exceeds the table.","tokens_in":25309,"tokens_out":6526,"would_cite":true,"duration_ms":64963,"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":"This paper proposes that distance-based OOD detection can be improved per input by shrinking the within-class covariance along the residual-space direction of each real-time feature, without retraining the model.","keywords":["out-of-distribution detection","Mahalanobis distance","dynamic covariance calibration","residual space projection","post-hoc OOD detection","information geometry","self-supervised models","covariance shrinkage"],"falsifier":"A controlled experiment on synthetic features would settle the central claim: train a Gaussian in-distribution model with an injected outlier cluster in direction $u$, then evaluate on OOD samples drawn from a direction $v$ orthogonal to $u$. If the dynamic score still beats static Mahalanobis, the outliers-align-with-novel-directions hypothesis is not the mechanism; if the advantage vanishes, the hypothesis is confirmed. Additionally, one could search for real feature spaces where some inputs give $p\\le 1$ or $s<0$, which would make the adjusted matrix non-positive-definite and produce invalid distance scores.","tokens_in":1965,"feed_emoji":"🎯","tokens_out":2098,"duration_ms":84534,"temperature":0.7,"pith_summary":"The paper claims that the geometry used by distance-based out-of-distribution (OOD) detection can be corrected at test time instead of being taken as a static object from training data. Its proposal is to watch each new input feature and shrink the within-class covariance along that feature's projection onto the residual space of the training distribution, making the Mahalanobis-like distance more sensitive to deviations in exactly that direction. The paper argues that training outlier features distort the covariance along directions that new OOD features also occupy, so a dynamic, per-input adjustment separates in- and out-of-distribution samples better than a static one. If true, OOD detection for any pre-trained feature extractor improves in a purely post-hoc manner, with no retraining of the model. The claim is supported by experiments across CIFAR and ImageNet-1k backbones, including the self-supervised DINO model.","feed_headline":"Per-input covariance tweak sharpens OOD detection","feed_subtitle":"Each new feature shrinks the geometry along its own residual direction, improving post-hoc detection.","key_machinery":"The central object is the dynamically adjusted precision matrix $M(f) = (\\Sigma_R - B^\\top a a^\\top B)^{-1}$, an online rank-one downdate of the within-class covariance before inversion. $B$ collects the residual-space eigenvectors of $\\Sigma_R$, and $a = B^\\top f$ projects the live feature onto that space. The Sherman-Morrison formula supplies the theoretical link: it shows how subtracting the outer product changes the inverse, and Theorem 4.2 converts the requirement that the adjusted form stay non-negative into the checkable inequalities $p>1$, $q>1$, $s\\ge 0$. The scoring function then takes the smallest class-mean distance under this per-input geometry.","core_discovery":"On its own terms, the paper's discovery is a score function that recalibrates the Mahalanobis geometry per test point. Starting from the within-class covariance matrix $\\Sigma_R$ of centered training features, the score for a real-time feature $f$ is $s(f) = -\\min_i \\sqrt{r_i^\\top(\\Sigma_R - B^\\top a a^\\top B)^{-1} r_i}$, where $r_i = f - \\mu_i$ is the deviation from class mean $i$, $B$ is the basis of the residual space (the eigenvectors of $\\Sigma_R$ belonging to its smallest eigenvalues), and $a = B^\\top f$ is the coefficient vector of $f$ in that residual space. Subtracting $B^\\top a a^\\top B$ removes the component of variance along $f$'s residual direction before the matrix is inverted, so the distance metric becomes locally more sensitive to novel directions while leaving the principal space untouched. A theorem gives conditions ($p>1$, $q>1$, $s\\ge 0$) under which the adjusted quadratic form is a valid, non-negative distance; the paper reports that these conditions hold on its ImageNet experiments. The paper reports that this score improves both AUROC and FPR95 over the compared post-hoc detectors on seven pre-trained models.","pith_inferences":["Editorial inference: the update is essentially a rank-one downdate of the precision matrix along the live sample's residual direction; one natural extension the paper does not develop is to accumulate several test-point downdates into a batch-level update, which would drift the geometry over a test stream.","Editorial inference: the central hypothesis that training outliers align with OOD directions could be probed directly by injecting synthetic outlier directions into Gaussian features; if OOD samples are drawn from a direction orthogonal to the injected outlier, the score's advantage over static Mahalanobis should shrink or vanish, and if it does not, the paper's explanation would need revision.","Editorial inference: the same dynamic-geometry idea should transfer to retrieval or open-set recognition tasks where each query defines a local direction along which the feature metric should be contracted.","Editorial inference: the OpenOOD near-OOD table shows the dynamic update helps RMDS but hurts plain Mahalanobis on some splits, so the practical claim is not that every distance improves, but that the geometry choice and the dynamic update interact; a pairing rule for which base distance to use is left open."],"forward_implications":["Post-hoc OOD detection can be improved on a frozen pre-trained feature extractor by adding a per-input covariance update, so the method composes with existing backbones without any training phase.","The residual-space restriction preserves the principal structure of the in-distribution data, which is why the score stays meaningful on weaker encoders such as ResNet-50 where static subspace methods struggle.","The dynamic adjustment applies not only to Mahalanobis distance but to any covariance-induced distance; the appendix shows it also helps Euclidean distance and RMDS in near-OOD settings.","Because the condition for a valid distance holds empirically when class clusters are compact and far from the origin, the method is expected to transfer to other well-separated feature spaces, including self-supervised ones like DINO."],"supporting_citations":[{"why":"Supplies the baseline Mahalanobis distance score whose static covariance the paper recalibrates.","marker":"Lee et al., 2018"},{"why":"Supplies the residual-space basis $B$ (eigenvectors of the smallest eigenvalues) and the claim that OOD features carry more energy in residual space, the geometric premise for restricting adjustments.","marker":"Wang et al., 2022"},{"why":"Supports the orthogonal and neural-collapse assumption that OOD features occupy the null space of ID features, justifying residual-space adjustment.","marker":"Ammar et al., 2024"},{"why":"The closest distance-based competitor; its sensitivity to class separability motivates the dynamic covariance design and is the main baseline beaten on CIFAR.","marker":"Liu & Qin, 2023"},{"why":"Shows in the appendix that the dynamic update transfers to another Mahalanobis variant in near-OOD detection.","marker":"Ren et al., 2021"},{"why":"Distance-based baseline whose feature-quality assumptions the paper compares against on both benchmarks.","marker":"Sun et al., 2022"},{"why":"Supplies the ImageNet OOD benchmark protocol and curated SUN, Places, and Texture subsets used in evaluation.","marker":"Huang & Li, 2021"},{"why":"Pre-trained self-supervised backbone used as one of the seven evaluation models, establishing the method works without labels.","marker":"Caron et al., 2021"}],"fun_headline_variants":["Dynamic covariance tuning boosts OOD detection","Per-test-point covariance update tightens OOD detection","Residual-space covariance adjustment refines OOD scores","Calibrating covariance per input improves OOD detection","Dynamic geometry correction enhances outlier detection"],"cache_read_input_tokens":27392,"weakest_assumption_plain":"The method assumes that the directions in which training features are distorted by outliers are the same directions in which new OOD test features lie, and that for every input the quantities $p>1$, $q>1$, $s\\ge 0$ hold so the adjusted covariance stays positive definite.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic covariance tuning boosts OOD detection","Per-test-point covariance update tightens OOD detection","Residual-space covariance adjustment refines OOD scores","Calibrating covariance per input improves OOD detection","Dynamic geometry correction enhances outlier detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000645,"raw_usage":{"total_tokens":3002,"prompt_tokens":1024,"completion_tokens":1978,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":640,"completion_tokens_details":{"reasoning_tokens":1909}},"tokens_in":640,"tokens_out":1978,"duration_ms":14576,"temperature":1.0,"reasoning_tokens":1909,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:49:16.841491+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment on synthetic features would settle the central claim: train a Gaussian in-distribution model with an injected outlier cluster in direction $u$, then evaluate on OOD samples drawn from a direction $v$ orthogonal to $u$. If the dynamic score still beats static Mahalanobis, the outliers-align-with-novel-directions hypothesis is not the mechanism; if the advantage vanishes, the hypothesis is confirmed. Additionally, one could search for real feature spaces where some inputs give $p\\le 1$ or $s<0$, which would make the adjusted matrix non-positive-definite and produce invalid distance scores.","supporting_citations":[{"cited_title":"Vim: Out-of-distribution with virtual-logit matching","cited_arxiv_id":null,"evidence_quote":"Supplies the residual-space basis $B$ (eigenvectors of the smallest eigenvalues) and the claim that OOD features carry more energy in residual space, the geometric premise for restricting adjustments."},{"cited_title":"B., Belkhir, N., Popescu, S., Manzanera, A., and Franchi, G","cited_arxiv_id":null,"evidence_quote":"Supports the orthogonal and neural-collapse assumption that OOD features occupy the null space of ID features, justifying residual-space adjustment."}],"review_version":1}