acceptodds
Under review as a conference paper at ICLR 2027

Diversity As a Paradigm for Out-of-Distribution Detection

Abstract

Out-of-distribution (OOD) detection is critical for the safe deployment of machine learning systems and for scientific discovery. Existing post-hoc detectors typically rely on model confidence scores or likelihood estimates in feature space, often under restrictive distributional assumptions. In this work, we introduce a new paradigm and formulate OOD detection from a diversity perspective. We propose Vendi Score Out-Of-Distribution detection (VSOOD), an OOD detector based on the Vendi Scores (VS), a family of similarity-based diversity metrics. VSOOD quantifies how much a test sample increases the VS of the in-distribution feature set, providing a principled notion of novelty that does not require density modeling. VSOOD is linear-time and naturally combines class-conditional (local) and dataset-level (global) novelty signals. Across multiple image classification benchmarks and network architectures, VSOOD achieves state-of-the-art OOD detection performance. Remarkably, VSOOD retains this performance when computed using only 1% of the training data, enabling deployment in memory- or access-constrained settings.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
OOD Detection, Vendi Scoring, Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_haMHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes VSOOD, a post-hoc out-of-distribution detector that measures how much adding a test feature increases the diversity of an in-distribution feature set as quantified by the Vendi Score. The method builds class-conditional feature statistics, uses a rank-1 approximation that retains the leading eigenvalue/eigenvector for each class, aggregates class-conditional novelty over the top-K predicted classes using probability weights, and optionally subtracts a global Vendi-based correction. Experiments on CIFAR-10, CIFAR-100, and ImageNet-1K with ResNet, ViT, and Swin architectures show competitive average performance, low inference overhead, and strong data efficiency, including results using only a small fraction of the training data.

Critical strengths
  1. Efficient and practical formulation. The rank-1 approximation reduces the class-conditional computation to O(CD), and the paper provides useful ablations showing that rank-1 performs very similarly to higher-rank variants on the reported ImageNet settings. The method is also attractive in low-data/access settings, where it remains effective with only a small fraction of the training data.
  2. Broad cross-architecture evaluation. The paper evaluates multiple post-hoc OOD paradigms on CIFAR-10/100 and ImageNet-1K using ResNet, ViT, and Swin models. VSOOD is especially competitive in the cross-architecture ImageNet average and provides a favorable accuracy/latency trade-off relative to more expensive feature-density baselines.
Critical weaknesses
  1. The novelty of the Vendi-based formulation is not sufficiently isolated from a much simpler principal-direction similarity detector. Under the proposed rank-1 approximation, for a fixed class the only test-sample-dependent term in the local score is alpha_c(x) = (u_c^T h(x))^2; the remaining quantities are class constants. Thus the local VSOOD score is a monotonic transformation of squared similarity to the leading class direction. The paper does not include a direct baseline that uses this simple principal-direction score with the same top-K probability weighting (and, if desired, the same global correction). This comparison is important for establishing that the gain comes from the Vendi/diversity formulation rather than from a simple first-principal-direction score plus aggregation. This concern is reinforced by Table 4, where a probability-weighted RMDS++ hybrid is already very competitive and is better than VSOOD in some settings. The paper should therefore moderate the 'new paradigm' framing unless this contribution is more clearly isolated.

  2. The empirical evidence is strong in aggregate but not consistently superior enough to fully support the broad state-of-the-art framing. On ImageNet-1K, VSOOD is clearly weaker than activation-shaping methods on ResNet-50 and is slightly below RMDS++ on ViT-B/16, while its main advantage appears in the cross-architecture average and on Swin-T. The CIFAR gains are comparatively modest. In addition, although the baseline suite contains several recent methods, it does not appear fully current for a 2027 submission (e.g., recent 2025-2026 post-hoc methods, which weakens the strongest SOTA claims. There is also some asymmetry in tuning effort: VSOOD receives a nontrivial grid search over (K, gamma, g), whereas KPCA uses recommended defaults because a full sweep is considered too expensive. Finally, the presentation could be substantially improved: Section 3 devotes too much main-text space to algebraic expansions and perturbation details. The core method would be easier to assess if the main text retained only the definition, final rank-1 score, aggregation, and algorithmic intuition, with most intermediate derivations moved to the appendix.

Questions for authors
  1. Could the authors add a simple principal-direction baseline based directly on alpha_c(x) = (u_c^T h(x))^2, using exactly the same top-K probability weighting and global-correction protocol as VSOOD? If such a baseline performs substantially worse, it would materially strengthen the claim that the Vendi formulation contributes beyond first-principal-direction similarity; if it performs similarly, the novelty claim should be reconsidered.

  2. Could the authors update the comparison with the most relevant recent post-hoc OOD methods (including representative 2025-2026 approaches) and provide a clearer architecture-wise analysis of where VSOOD helps or fails? In particular, why does VSOOD underperform the best activation-shaping methods on ResNet-50 but become strongest on Swin-T, and does this behavior persist under a matched hyperparameter-tuning budget across methods?

2 comments
AuthorsOpenReview ↗

Response to Reviewer haMH

We thank the reviewer for their feedback.

Contribution of VSOOD Beyond Principal-Direction Similarity

Following the reviewer's suggestion, we introduce a new baseline to isolate the contribution of the Vendi novelty formulation. We call this baseline PCA-Align for simplicity. PCA-Align replaces the class-conditional Vendi novelty with 1−αc(x)1-\alpha_c(x), where αc(x)=(uc⊤h(x))2\alpha_c(x)=(u_c^\top h(x))^2 measures squared alignment with the leading class eigenvector. We make the analogous substitution for the global correction: SPCA-Align(x)=∑c∈TK(x)pc(x)γ(1−αc(x))−g(1−αglobal(x))S_{\text{PCA-Align}}(x) = \sum_{c \in T_K(x)} p_c(x)^{\gamma}(1-\alpha_c(x)) - g(1-\alpha_{\text{global}}(x))

PCA-Align retains VSOOD's top-K class selection, probability weighting, global-correction structure, and hyperparameter-selection protocol, replacing only the Vendi novelty terms with their alignment-based counterparts. Unlike incremental Vendi novelty, these alignment scores do not require normalization by training-set size. Thus, the comparison directly tests whether principal-direction alignment is sufficient to explain VSOOD's performance.

We also add Centroid-Align, which replaces the Vendi novelty terms with cosine distance to class centroids. These baselines are described in Section 4.2 and results are reported in revised Table 4 in the main text.

To address the concern about tuning fairness, we reimplemented all KPCA-based experiments, including standalone KPCA and the KPCA+GEN hybrid, using the updated TPAMI 2026 method [1] and tuning kernel bandwidth and PCA dimension for each setting. The resulting average AUROC comparisons are:

DatasetArchitectureHybrid RMDS++KPCA+GENPCA-AlignCentroid-AlignVSOOD
CIFAR-10ResNet-1891.2192.6492.9093.1793.13
CIFAR-100ResNet-1881.5580.4781.9882.0282.29
ImageNet-1KResNet-5088.1087.2488.4188.5389.26
ImageNet-1KViT-B/1688.3586.8886.1386.4987.48
ImageNet-1KSwin-T88.5486.3086.2587.7588.79
ImageNet-1KDenseNet-20186.7984.6888.2988.3388.75
ImageNet-1KMaxViT-T87.7087.3287.3287.5487.71
ImageNet-1KMobileNetV3-Large82.8378.0181.9082.0584.97
ImageNet-1KInception-v387.0083.5088.9788.6688.71

VSOOD outperforms PCA-Align, Centroid-Align, and Hybrid RMDS++ individually on 8 of 9 settings, and KPCA+GEN on all 9 settings. It achieves the highest AUROC on six settings and ranks second on the remaining three. Averaged across seven ImageNet architectures, VSOOD improves AUROC by 1.20 points over PCA-Align and 0.90 points over Centroid-Align.

Mathematical Distinction from PCA Similarity

We agree with the reviewer that, for a fixed class, the rank-1 Vendi novelty is a monotonic function of the squared principal-direction alignment. However, this does not imply that VSOOD and PCA-Align produce equivalent OOD scores, as they transform and aggregate these class-conditional similarities differently.

As shown in Appendix A.2, the leading-order form of our novelty score is NcΔc(x)≈2(1−αc(x)/λc)N_c\Delta_c(x)\approx 2(1-\alpha_c(x)/\lambda_c), compared with the squared projection residual 1−αc(x)1-\alpha_c(x) used by PCA-Align. The normalization by λc\lambda_c has a natural interpretation: it is the average squared alignment of the training samples with the class's leading eigenvector, 1Nc∑i:yi=c(uc⊤h(xi))2=uc⊤ρcuc=λc.\frac{1}{N_c}\sum_{i:y_i=c}(u_c^\top h(x_i))^2 =u_c^\top\rho_cu_c=\lambda_c.

Thus, VSOOD measures a sample's alignment relative to what is typical for its class. Since the eigenvalue λc\lambda_c varies across classes, this distinction matters when combining class-conditional novelty scores through probability-weighted aggregation.

References

[1] Fang, Kun, et al. "Kernel PCA for out-of-Distribution Detection: Non-Linear Kernel Selection and Approximation." IEEE Transactions on Pattern Analysis and Machine Intelligence (2026).

AuthorsOpenReview ↗

Response to Reviewer haMH (continued)

Expanded Experimental Evaluation and Recent Baselines

To strengthen the empirical evaluation, we added four ImageNet architectures: DenseNet-201, MaxViT-T, MobileNetV3-Large, and Inception-v3. These extend our evaluation from three to seven pretrained classifiers, spanning convolutional, transformer, and hybrid architectures.

We also added five recent post-hoc OOD detection methods: LogitGAP [1], RAS [2], CADRef [3], FOLD [4], and BNP4OOD [5]. These span logit-based detection, activation shaping, feature geometry, and Bayesian nonparametric modeling. We additionally updated the KPCA implementation to the 2026 TPAMI version [6], as requested by another reviewer. The kPCA results now sweep over 64 configurations of kernel bandwidth and PCA dimension.

The expanded results are provided in revised Table 2, with architecture-specific results in Tables 7–13. Averaged across all seven ImageNet architectures, VSOOD achieves the strongest performance on Near-OOD and Far-OOD sets.

MethodNear-OOD AUROC AVGFar-OOD AUROC AVG
MSP75.7885.31
ASH57.6962.16
SCALE67.5777.20
ADASCALE74.7490.02
GEN77.3689.60
IODIN75.8386.30
KNN68.8886.85
FDBD72.7888.86
NCI74.2590.04
KPCA62.3582.72
RMDS++78.5090.45
Fold74.5087.08
LogitGAP77.4989.07
BNP4OOD76.5986.59
RAS72.8188.61
CADRef74.7591.51
VSOOD80.3393.04

Compared with the strongest competing methods, VSOOD improves average near-OOD AUROC by 1.83 points and far-OOD AUROC by 1.53 points. It also achieves the lowest average FPR@95 (41.61 versus 46.73 for the next-best method). Across the seven architectures, VSOOD achieves the highest average AUROC on four and ranks among the top three on all seven. These additional experiments show that the advantages are not confined to Swin-T or a particular family of pretrained classifiers.

Architecture-Specific Performance

We also investigated why activation-shaping methods perform particularly well on ResNet-50 but degrade on architectures such as Swin-T.

Prior work by [7] motivates activation scaling through the observation that ID samples often have larger positive activations than OOD samples. We examined this relationship across all seven ImageNet models and added the results to Table 14.

ImageNet-1K ModelID MeanOOD MeanID/OOD Ratio
ResNet-500.46700.37971.230
ViT-B/160.22170.25710.862
Swin-T0.18140.22230.816
DenseNet-2010.25680.22701.131
MaxViT-T0.12950.14960.866
MobileNetV3-Large0.13520.14930.905
Inception-v30.25150.28540.881

The activation-shaping methods ADASCALE and SCALE provide strong performance on ResNet (Table 8) and DenseNet (Table 10) - the two models for which the assumption in [7] is true. VSOOD outperforms ADASCALE on all other baselines. We have added this new analysis and discussion to our paper in Table 14 and Section 4.1.

Improved Presentation

We have incorporated your feedback in Section 3. We moved most intermediate derivations to the appendix, removed redundant equations, and re-phrased certain parts to improve its overall readability.

References

[1] Liang, Jiachen, et al. "Revisiting logit distributions for reliable out-of-distribution detection." Advances in Neural Information Processing Systems (2025).

[2] Guglielmo, Gianluca, and Marc Masana. "Ranked Activation Shift for Post-hoc Out-of-Distribution Detection." European Conference on Computer Vision. (2026).

[3] Ling, Zhiwei, et al. "Cadref: Robust out-of-distribution detection via class-aware decoupled relative feature leveraging." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. (2025).

[4] Park, Seonghwan, et al. "Exploiting Local Flatness for Efficient Out-of-Distribution Detection." European Conference on Computer Vision. (2026).

[5] Linderman, Randolph W., et al. "A bayesian nonparametric perspective on mahalanobis distance for out of distribution detection." Transactions on Machine Learning Research (2026).

[6] Fang, Kun, et al. "Kernel PCA for out-of-Distribution Detection: Non-Linear Kernel Selection and Approximation." IEEE Transactions on Pattern Analysis and Machine Intelligence (2026).

[7] Xu, Kai et. al, “Scaling for training time and post-hoc out-of-distribution detection enhancement.” arXiv:2310.00227 (2023).

Reviewer_DfKKOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This work introduces the Vendi Score (VS) into the Out-Of-Distribution (OOD) detection task, and proposes a novel method namely VSOOD. The key of VSOOD is to compare the log-ratio increase in VS from adding a test sample to the training data. For the local version of VSOOD, this log-ratio VS increase is calculated per class and is then combined with predictive probabilities, leading to the local score. For the global version of VSOOD, this log-ratio VS increase is calculated based on all training samples. The final detection score is set as the differences between the local score and the scaled global score. Experiments validate the effectiveness of VSOOD.

Critical strengths
  1. It is interesting to introduce the Vendi Score concept into the OOD detection community.
  2. The authors show how to accelerate calculations for the VS-based detection score.
  3. Experiments cover the large-scale ImageNet-1K dataset.
Critical weaknesses

Note that this is an OOD detection task. Therefore, the effectiveness of Vendi Score largely relies on whether the similarity matrix (or kernel matrix) correctly characterizes the relationships between In-Distribution (ID) data and OOD data. In this work, VSOOD uses a cosine kernel (line 150). However, VSOOD does not sufficiently justify how the cosine kernel benefits ID-OOD separation, and VSOOD actually adopts the cosine kernel largely due to computational efficiency considerations. In this regard, I came across a journal version [a] of the compared baseline kernel PCA approach, which provides a simple justification for the selected cosine kernel. The authors are encouraged to include [a] in the references for an up-to-date literature review, and to provide a justification for the choice of the cosine kernel for the OOD detection task, particularly under this VS framework (if possible).

[a] Kernel PCA for Out-of-Distribution Detection: Non-Linear Kernel Selection and Approximation. IEEE TPAMI 2026.

Questions for authors
  1. As summarized in the weakness above, could the authors provide a justification for the choice of the cosine kernel, particularly under this VS framework, for the OOD detection task?

  2. The authors use a validation set to search the three hyper-parameters. Are there any heuristic strategies or recommended initializations to help determine these hyper-parameters?

1 comment
AuthorsOpenReview ↗

Response to Reviewer DfKK

We thank the reviewer for their feedback.

Updated kPCA Reference

We have updated our manuscript to use the latest implementation of the kernel PCA approach, and have updated the citation accordingly. We reimplemented the updated approach and revised all KPCA-based results, including KPCA+GEN, accordingly. To ensure a fairer comparison, we also tune its associated hyperparameters, including kernel bandwidth and PCA dimension instead of the fixed defaults used in the original submission.

Exploring the Choice of Kernel

Our choice of cosine similarity is motivated by both its suitability for measuring diversity and its computational efficiency. The cosine kernel measures diversity in terms of the angular relationships between learned representations, while removing the influence of feature magnitudes. This is desirable because feature norms can vary substantially across pretrained classifiers and may not reliably characterize ID-OOD separation, as also investigated by [1]. To further investigate this choice, we added a kernel ablation in Appendix C (Table 6), evaluating RBF, Laplace, cosine-RBF, and cosine-Laplace kernels in addition to the original cosine kernel. The average AUROC results are:

DatasetArchitectureCosineRBFCos-RBFLaplaceCos-Laplace
CIFAR-10ResNet-1893.1391.3792.3891.4591.57
CIFAR-100ResNet-1882.2982.0981.7682.1981.80
ImageNetResNet-5089.2685.8889.2985.8888.65
ImageNetViT-B/1687.4887.3388.6387.9188.74
ImageNetSwin-T88.7986.1288.2986.9588.40
ImageNetDenseNet-20188.7585.3288.7885.5488.43
ImageNetMaxViT-T87.7187.3988.3888.0489.10
ImageNetMobileNetV3-L84.9776.8884.8078.5885.35
ImageNetInception-v388.7185.4888.4785.9388.19

Cosine achieves the best performance on four of nine settings. However, fused kernels can provide further improvements on particular architectures. For example, cosine-Laplace improves AUROC from 87.48 to 88.74 on ViT-B/16 and from 87.71 to 89.10 on MaxViT-T. These alternative kernels can be approximated using Random Fourier Features, allowing VSOOD to retain its efficiency. Overall, the results support cosine as a strong, simple, and computationally efficient default, while demonstrating that VSOOD is not restricted to cosine similarity and can benefit from richer kernels. We have incorporated the kernel-selection discussion into Section 4.1 and Appendix C (Table 6).

Hyperparameter Selection

Although we recommend the validation-based procedure described in Section 4, we investigated whether simpler fixed configurations or heuristics based entirely on ID training data could provide competitive performance.

Fixed Configuration. As a practical default for ImageNet-1K, we recommend K=50, γ=0.5\gamma=0.5, and g=1. This configuration achieves an average AUROC of 87.36 across seven architectures, compared with 87.95 using validation-selected hyperparameters, a difference of only 0.59 AUROC points.

ID-Only Heuristics. We also tested two strategies that select hyperparameters without synthetic or real OOD validation samples:

  • Safe Tail Support: Select K based on the 90th percentile of the number of classes whose softmax probabilities are at least 10-3 times the maximum class probability on ID samples. We fix γ=0.5\gamma=0.5 and g=1.
  • Feature–Logit Consistency: Construct synthetic feature–logit mismatches by pairing an ID feature vector with logits from an example of a different class. We select K, γ\gamma, and g to maximize the separation between matched and mismatched pairs.

The resulting average AUROC values are:

ConfigurationResNet-50ViT-B/16Swin-TDenseNet-201MaxViT-TMobileNetV3-LInception-v3Average
Safe Tail Support87.8486.5288.4288.0287.3884.6687.7687.23
Feature–Logit Consistency88.3886.6288.7688.5486.3983.1788.4387.19
Fixed (K=50,gamma=0.5,g=1)87.8786.7788.4288.0187.7184.9787.7687.36
Validation-Tuned VSOOD89.2687.4888.7988.7587.7184.9788.7187.95

Both ID-only heuristics achieve competitive average performance, although neither consistently outperforms the fixed configuration. These experiments suggest that VSOOD can achieve strong performance without a dedicated hyperparameter search.

Nevertheless, we recommend the original validation-based procedure. It achieves the strongest average performance, and its computational cost is modest: the validation set consists only of ID samples and automatically generated Gaussian-noise images, requiring no access to real OOD data and no additional classifier training.

References

[1] Müller, Maximilian, and Matthias Hein. "Mahalanobis++: Improving OOD detection via feature normalization." arXiv:2505.18032 (2025).

Reviewer_KTpcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper tackles OOD detection through diversity in feature space (compared to previous classes) via the (generalized) Vendi score (Renyi entropy of eigenvalues) of the cosine similarity kernel matrix. Basically they test whether a new point would change the Vendi score/feature diversity much (first class-conditional, then weighted via class probabilities).  The rank-1 approximation they actually implement for computational simplicity is a combination of cosine-similarity with PC style methods and weighting with predictive probabilities - both of which have appeared before.

Critical strengths

The writing and math is pretty easy to follow (albeit with unnecessarily many equations/steps in the main text) and the flow is reasonable.

Critical weaknesses

The paper does mention prior work using feature embeddings but does not really say how the Vendi score provides an additional advantage over those (both mathematically speaking and

In a similar light of the missing intuition, the experiments are not convincingly conveying that the score is actually more informative than say just measuring similarity with class-means (compared to the reference of known in class points)? Further, Vendi score motivation is to that it reflects how many truly different points there are in the sample (diversity in the data). When the authors do a rank-1 approx (and indeed they show experimentally that it doesn't make a huge difference to increase the rank) the benefits of the Vendi score gets lost, and  it seems to me to contradict the whole motivation behind the Vendi-score; the fact that rank-1 seems sufficient  suggests to me that perhaps the extension beyond PCA-style methods (using just the average feature vector of a class) might not be too large. 

Indeed experimentally, in the relevant near-ood benchmarks, very related methods that are spiritually close to the rank-1 approximated Vendi score (which, looking at A.1. is just 'how close is the test sample feature vector to the principle component/mean of that class - weighted by the classes) in the baseline comparisons do similarly well (NCI, KPCA or RMDS++ - there are also many other very related methods that weren't cited). The experimental difference in performance is imv not large enough to really publish a paper on this method in my view.

Hence, it seems to me having a lack of truly novelty and it's unclear what this paper adds to the field.

Questions for authors

There's nothing concrete the authors could do that would lead to a score raise to accept.

2 comments
AuthorsOpenReview ↗

Response to Reviewer KTpc

We thank the reviewer for their comments.

What Does the Vendi Score Contribute Beyond PCA and Class-Mean Similarity?

To directly address the reviewer's concern about the utility of the Vendi contribution, we add two new controlled baselines, PCA-Align and Centroid-Align. Both retain VSOOD's top-K probability-weighted aggregation, optional global correction, and validation-selection procedure, while replacing the Vendi-derived novelty function. PCA-Align replaces the class-conditional Vendi novelty with 1−αc(x)1-\alpha_c(x), where $\alpha_c(x) = (u^T_c h(x))^2 $ measures squared alignment with the leading eigenvector of the class-conditional matrix ρc\rho_c. We make the analogous substitution for the global correction, yielding SPCA-Align(x)=∑c∈TK(x)pc(x)γ(1−αc(x))−g(1−αglobal(x))S_{\text{PCA-Align}}(x) = \sum_{c \in T_K(x)} p_c(x)^{\gamma}(1-\alpha_c(x)) - g(1-\alpha_{\text{global}}(x))

Centroid-Align instead uses (1−cos⁡(h(x),μc))(1-\cos(h(x),\mu_c)), directly testing whether similarity to class means is sufficient.

We select KK, γ\gamma, and gg using the same validation protocol as VSOOD. The resulting average AUROC comparisons are:

DatasetArchitecturePCA-AlignCentroid-AlignVSOOD
CIFAR-10ResNet-1892.9093.1793.13
CIFAR-100ResNet-1881.9882.0282.29
ImageNet-1KResNet-5088.4188.5389.26
ImageNet-1KViT-B/1686.1386.4987.48
ImageNet-1KSwin-T86.2587.7588.79
ImageNet-1KDenseNet-20188.2988.3388.75
ImageNet-1KMaxViT-T87.3287.5487.71
ImageNet-1KMobileNetV3-Large81.9082.0584.97
ImageNet-1KInception-v388.9788.6688.71

VSOOD outperforms each alternative on 8 of the 9 benchmarks, and outperforms both simultaneously on 7 of 9. These comparisons establish that VSOOD's performance cannot be explained solely by measuing principal direction alignment or similarity to class centroids. The Vendi-derived novelty formulation provides a consistently more effective score. These results are now included in Table 4 in the main text.

Mathematical Distinction from PCA Similarity

Although VSOOD and rank-1 PCA use the same leading eigenvector, they define novelty differently. PCA-Align measures the squared projection residual along the leading class eigendirection, whereas VSOOD measures the change in a class's diversity when a test sample is added. As shown in Appendix A.2, the leading-order form of our novelty score is NcΔc(x)≈2(1−αc(x)/λc)N_c\Delta_c(x)\approx 2(1-\alpha_c(x)/\lambda_c), compared with the squared projection residual 1−αc(x)1-\alpha_c(x) used by PCA-Align. The normalization by λc\lambda_c has a natural interpretation: it is the average squared alignment of the training samples with the class's leading eigenvector, 1Nc∑i:yi=c(uc⊤h(xi))2=uc⊤ρcuc=λc.\frac{1}{N_c}\sum_{i:y_i=c}(u_c^\top h(x_i))^2 =u_c^\top\rho_cu_c=\lambda_c.

Thus, to leading order, VSOOD measures a sample's alignment relative to what is typical for its class, whereas PCA measures absolute misalignment without accounting for differences in class spectral concentration.

AuthorsOpenReview ↗

Response to Reviewer KTpc (continued)

Expanded Experimental Coverage and Baselines

We have substantially expanded our experimental evaluation to address concerns regarding the generality of the empirical improvements and the coverage of related methods.

First, we added five recent OOD detection baselines: LogitGAP [1], RAS [2], CADRef [3], FOLD [4], and BNP4OOD [5]. These span logit-based detection, activation shaping, feature geometry, and Bayesian nonparametric modeling. We additionally updated the KPCA implementation to the 2026 TPAMI version, as requested by another reviewer. In total, our evaluation now includes 16 baseline algorithms.

Second, we expanded ImageNet-1K evaluation from three to seven pretrained architectures, adding DenseNet-201, MaxViT-T, MobileNetV3-Large, and Inception-v3. This substantially broadens the architectural diversity of our evaluation, including convolutional, transformer, and hybrid models. The complete results appear in revised Tables 2 and 7–13. VSOOD achieves the strongest average performance across all seven ImageNet architectures:

MethodNear-OOD AUROC AVGFar-OOD AUROC AVG
MSP75.7885.31
ASH57.6962.16
SCALE67.5777.20
ADASCALE74.7490.02
GEN77.3689.60
IODIN75.8386.30
KNN68.8886.85
FDBD72.7888.86
NCI74.2590.04
KPCA62.3582.72
RMDS++78.5090.45
Fold74.5087.08
LogitGAP77.4989.07
BNP4OOD76.5986.59
RAS72.8188.61
CADRef74.7591.51
VSOOD80.3393.04

VSOOD outperforms the strongest competing method by 1.83 AUROC points on near-OOD and 1.53 points on far-OOD. VSOOD also achieves the best seven-model average FPR@95 of 41.61, compared with 46.73 for the next-best method. Across the seven architectures, VSOOD achieves the highest average AUROC on four and ranks among the top three on all seven.

References

[1] Liang, Jiachen, et al. "Revisiting logit distributions for reliable out-of-distribution detection." Advances in Neural Information Processing Systems (2025).

[2] Guglielmo, Gianluca, and Marc Masana. "Ranked Activation Shift for Post-hoc Out-of-Distribution Detection." European Conference on Computer Vision. (2026).

[3] Ling, Zhiwei, et al. "Cadref: Robust out-of-distribution detection via class-aware decoupled relative feature leveraging." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. (2025).

[4] Park, Seonghwan, et al. "Exploiting Local Flatness for Efficient Out-of-Distribution Detection." European Conference on Computer Vision. (2026).

[5] Linderman, Randolph W., et al. "A bayesian nonparametric perspective on mahalanobis distance for out of distribution detection." Transactions on Machine Learning Research (2026).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Summary of Revisions and Rebuttal Responses

Based on the reviewer’s feedback, we have included new experiments to provide stronger evidence for the value of our proposed method (VSOOD). We have uploaded our revised manuscript and summarize the key changes below.

Updated Baselines

We have incorporated 5 new baselines from 2025 and 2026 papers. These methods cover logit-only (LogitGAP [1]), activation-shaping (RAS [2]), and feature-geometry methods (BNP4OOD [3], FOLD [4], CADRef [5]). We also reimplemented both KPCA and KPCA+GEN using the updated TPAMI 2026 approach [6], including hyperparameter tuning for a fairer comparison. On ImageNet, VSOOD reduces the False Positive Rate@95 by an average of over 10% compared to the strongest baseline on ImageNet (Table 2), without retaining low latency overhead (Table 3). The updated aggregate results are reported in Table 2, with implementation details for the new baselines in Appendix B.

Additional Architectures

To assess whether VSOOD's improvements generalize across architectures, we expanded our ImageNet evaluation from three to seven classifiers, adding DenseNet-201, MaxViT-T, MobileNetV3-Large, and Inception-v3. VSOOD achieves the highest average AUROC on four of seven ImageNet architectures and ranks among the top three on all seven. The new results are provided in Tables 10–13, with aggregate results in Table 2. We also investigated the architecture-dependent behavior of activation-shaping methods such as ADASCALE. Our analysis of ID and OOD activation statistics offers a possible explanation for their greater performance variability across architectures (Section 4.1 and Appendix Table 14).

Comparison against PCA

To directly address concerns about the contribution of the Vendi formulation, we introduced PCA-Align, a controlled baseline that replaces Vendi novelty with principal-direction alignment while retaining the same probability weighting, top-K selection, and global-correction framework.

VSOOD outperforms PCA-Align on 8 of 9 evaluated settings. This provides strong evidence that VSOOD’s improvement cannot be explained solely by principal-direction similarity and probability-weighted aggregation. We report these results alongside other hybrid detectors in Table 4 and clarify the mathematical distinction in our rebuttal.

Kernel Selection

We added an ablation comparing the cosine kernel with RBF, Laplace, and fused cosine-based kernels (Appendix Table 6). The results support cosine as a strong and computationally efficient default, while showing that alternative kernels can improve performance on certain architectures.

Improved Methods Section

We incorporated reviewer feedback and shortened Section 3, retaining the main novelty definition, rank-1 approximation, and aggregation mechanism while moving intermediate derivations to the appendix. This improves the readability of our methods section without losing any mathematical precision.

References

[1] Liang, Jiachen, et al. "Revisiting logit distributions for reliable out-of-distribution detection." Advances in Neural Information Processing Systems (2025).

[2] Guglielmo, Gianluca, and Marc Masana. "Ranked Activation Shift for Post-hoc Out-of-Distribution Detection." European Conference on Computer Vision. (2026).

[3] Linderman, Randolph W., et al. "A bayesian nonparametric perspective on mahalanobis distance for out of distribution detection." Transactions on Machine Learning Research (2026).

[4] Park, Seonghwan, et al. "Exploiting Local Flatness for Efficient Out-of-Distribution Detection." European Conference on Computer Vision. (2026).

[5] Ling, Zhiwei, et al. "Cadref: Robust out-of-distribution detection via class-aware decoupled relative feature leveraging." 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition. (2025).

[6] Fang, Kun, et al. "Kernel PCA for out-of-Distribution Detection: Non-Linear Kernel Selection and Approximation." IEEE Transactions on Pattern Analysis and Machine Intelligence (2026).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.