acceptodds
Under review as a conference paper at ICLR 2027

LSF: Layer Selection and Fusion for Post-Hoc Out-of-Distribution Detection

Abstract

Distance-based out-of-distribution detection typically relies on the penultimate representation, whose optimization for ID class discrimination may suppress lower-level variation useful for detecting distribution shifts. We introduce LSF (Layer Selection and Fusion for OOD detection), a post-hoc framework that exploits complementary information across network layers. LSF selects an intermediate layer with the largest decrease in entropy density between consecutive layers and concatenates it with the penultimate representation to preserve their distinct feature spaces. Across ViT, Swin, and ConvNeXt architectures, LSF improves OOD detection across a range of evaluated settings, including near- and far-OOD and covariate shifts.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
OOD detection, Post-Hoc, Layer Selection, Layer Aggregation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_orhROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Distance-based OOD detection (e.g., Mahalanobis-family methods) almost exclusively uses penultimate-layer features, which are optimized for ID class discrimination and may suppress variations useful for detecting distribution shifts. This paper proposes LSF, which exploits complementary information from an intermediate layer to improve OOD separation.

Critical strengths
  1. Clean overall design: the layer-selection criterion is computable from ID data alone, requires no OOD data and introduces no layer-count hyperparameter; concatenation-based fusion naturally supports layers of unequal width, avoiding the restriction of X-Mahalanobis that additive fusion requires equal-dimensional features projected into a shared coordinate space.
  2. Minimal deployment overhead: only +3 KB memory and +0.0011 ms latency over single-layer Maha++, far below the multilayer baseline (X-Maha: 0.0464 ms), with better performance.
Critical weaknesses
  1. Evaluation-definition issue for covariate shift; the main gains may come from "low-level shift detection" rather than semantic uncertainty estimation. As the authors state in sec.4.1, ImageNet-C/R/V2/ES "retain the semantic classes of the ID data" (semantically ID), yet they are actually evaluated as OOD. This means the paper's main improvement (on covariate-shifted data) comes from rejecting more semantically unchanged same-class samples, which is not the OpenOOD setting. The slightly worse results on ImageNet-V2 (a freshly resampled, same-distribution test set) support this reading: no better performance when there is no low-level shift. The method's effectiveness likely stems from intermediate-layer features encoding priors on the background/low-level statistics of ID data, rather than achieving the semantic uncertainty estimation expected of OOD detection.
  2. Insufficient experiments on convolutional backbones: no ResNet; and if I read correctly, the full ConvNeXt-T results (Tables 11/12) are ImageNet-LT only, while under ImageNet-1K calibration ConvNeXt-T appears only in the Table 4 ablation — the results are incomplete. These should be supplemented.
  3. No mechanistic analysis of the source of the gains, nor of the advantage under LT calibration: as in Weakness 1, the gains concentrate on covariate shift/Textures with minimal semantic near-OOD gains, and the link between semantic shift and the intermediate-layer channel is not analyzed. Moreover, many experiments use ImageNet-LT calibration, yet the source of the method's advantage on such data is not analyzed either.
Questions for authors
  1. See Critical Weaknesses.
  2. Transparency of layer selection: which intermediate layer is actually selected (l*) for each backbone, and is it the same layer under ImageNet-1K vs. ImageNet-LT calibration? What mechanism governs this?
  3. KNN score inversion under LT calibration: in Table 8, KNN's Near-OOD AUROC is 38.16/40.83 (mean 48.63 on ConvNeXt-T, below 50 even for far-OOD), i.e., OOD samples receive higher ID scores. Negating the score restores AUROC > 50 — the information is present, the direction is wrong. Does this arise from the collapse of neighborhood sampling on the ID side (especially tail classes) for non-parametric density proxies under long-tailed calibration? This is quite interesting.
  4. Baseline reproduction details: please specify the exact hyperparameters and reproduction details of all compared methods.
2 comments

We thank the reviewer. We clarify shift detection, add convolutional results, and extend long-tail analyses. All references below use the revised manuscript.

W1. Semantic OOD versus covariate-shifted ID. Section 4.1 (Experimental Setup) explicitly distinguishes near/far semantic OOD (unseen classes) from covariate shifts retaining ImageNet classes. ImageNet-C/R/ES assess appearance/domain changes; ImageNet-V2 is identified as a close-distribution resampling control, not treated as equivalent to severe corruption. Sections 4.2–4.3, Tables 1–3, report semantic and covariate-shifted groups separately. We do not claim that greater detection of fog, defocus, or exposure changes demonstrates greater semantic uncertainty detection. Such monitoring can instead inform adaptation or recalibration. Figure 1 and Appendix C.4, Figures 6–7, illustrate why intermediate representations may respond differently to semantic and appearance changes.

W2. Convolutional-backbone coverage. Appendix C.3.2, Tables 14–15, now reports ConvNeXt-T under ImageNet-1K calibration: versus penultimate Maha++, LSF-C lowers mean semantic-OOD FPR95 44.13% → 40.56% and covariate-shift FPR95 70.54% → 64.45%. ConvNeXt-T/ImageNet-LT is in Tables 12–13; Swin-T/ImageNet-LT is in Tables 16–17. Appendix C.3.4, Tables 18–19, adds ResNet-50 (resnet50.tv2_in1k). With standard Mahalanobis, LSF-B lowers semantic/covariate FPR95 49.58/66.36% → 48.71/63.65% (AUROC 86.51/75.80% → 86.97/77.19%). With Maha++, LSF-C improves covariate FPR95 67.74% → 63.64% but worsens semantic FPR95 52.06% → 56.31%. The discussion in Appendix C.3.4 notes that the original Maha++ paper also found ResNet-50 sensitive to ℓ2\ell_2 normalization; this is context, not proof of the cause of the fused result. We explicitly report this limitation.

W3. Mechanism and long-tailed behavior. Section 3.3, Proposition 1(a), Eq. 11, decomposes the intermediate block's contribution to fixed-class expected Mahalanobis separation into residual mean and covariance effects; Proposition 1(b–c), Eqs. 12–13, gives a conditional spectral explanation of sensitivity and relates a consecutive entropy-density drop to a one-step threshold change under stated assumptions. Appendix A.2–A.4 gives proofs, conditional-KL examples, and favorable/unfavorable shifts. These results motivate, but do not guarantee, better FPR95. Section 4.4, Tables 4–6, studies selection, feature redundancy, and fusion under ImageNet-1K; Appendix C.6.1, Tables 25–27, adds candidate-layer sweeps, LT selection criteria, and LT redundancy. In Table 26, ViT-B/16 LSF-C (B4) lowers near/far/covariate FPR95 58.85/25.73/70.60% → 55.99/16.21/59.25% versus penultimate-only; ConvNeXt-T LSF-C (S1B0) records 64.21/26.84/66.30% versus 64.22/34.19/71.72%. Table 27 shows that adding a third layer raises mean FPR95 from 44.18% to 67.74%, alongside increased redundancy. This supports selective fusion, not a causal explanation of imbalance.

Q1. Which layers are selected under the two calibrations? The new Table 24 (Appendix C.6.1) answers this directly: ViT-B/16 B4; CLIP ViT-B/16 B6; Swin-T S2B0; ConvNeXt-T S1B0 under both ImageNet-1K and ImageNet-LT. Table 4 identifies ViT-B/16 and ConvNeXt-T selections for ImageNet-1K; Table 26 gives the LT comparison. This is empirical stability of the selected indices, not a claim that the covariance spectra are invariant to class frequencies. ResNet-50 is evaluated separately in Appendix C.3.4 and is not included in the four-backbone cross-calibration table.

Q2. KNN score inversion under ImageNet-LT. Appendix C.3.2 (KNN under long-tailed calibration) describes sparse tail-class neighborhoods and the change from k=1000k=1000 to dataset-scaled k=90k=90; Appendix D (KNN) specifies the two calibration settings. The updated KNN results appear in Appendix C.1, Tables 7, 8 and 9, and for ConvNeXt-T in Tables 12–13. Our additional neighborhood-size sensitivity check finds mean semantic-OOD AUROC 42.70% at k=1000k=1000, 67.24% at k=90k=90, and 80.88% at k=5k=5, consistent with the reviewer's hypothesis. This separate sweep is not presented as a numbered table in the manuscript. It does not establish a unique cause of inversion.

Q3. Baseline reproducibility. Section 4.1 names the benchmark, fixed backbones, detector families, metrics, and ID-only protocol. Appendix D (Implementation Details) specifies baseline scoring formulas and settings, including KNN k=1000/90k=1000/90, NNGuide k=10k=10, and RS-Maha's ID-only radial-exponent search. Appendix E (Reproduction Details) describes calibration, feature extraction, compute resources, and release materials; Table 32 maps model abbreviations to checkpoints. No baseline hyperparameter is tuned on OOD data or the ID test split.

We thank all reviewers for their time and thoughtful feedback!

In response to Reviewer p9sx's latest comments, we have further revised Section 3.3 and Appendix A.1–A.3 to strengthen the theoretical connection between entropy-density drops and potential OOD sensitivity, focusing on spectral sensitivity and complementary information. All references in our responses are to the latest revised manuscript.

As the discussion period draws to a close, we would greatly appreciate any further comments or questions from the reviewers, particularly regarding whether our revisions have addressed their concerns.

Reviewer_p9sxOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes LSF for selecting an intermediate activation using an entropy criterion. The selected activation is then concatenated with the penultimate representation and Mahalanobis scoring is applied for OOD detection. The approach achieves competitive performance across multiple architectures and distribution shifts.

Critical strengths
  • The paper demonstrates a successful use of intermediate features for OOD detection.
  • The authors test LSF on the OpenOOD benchmark, which provides sufficient breadth and stress tests for evaluating the method's performance. I also acknowledge that LSF achieves competitive performance when combined with Mahalanobis scoring.
Critical weaknesses
  • My primary concern is about the framing. The authors correctly acknowledge that LSF is an intermediate layer selection mechanism. It is distinguished from the prior work in how the selected activations are combined with the penultimate representation. In my opinion, the paper moves away from this perspectiev starting from Section 3.3 (also throughout the experimental section), the authors increasingly treat LSF combined with Mahalanobis scoring as an independent detector. I do not believe this is the correct way to present the contribution, as it takes the focus away from the main idea in terms of novelty.
  • The paper lacks scientific insight into why proposed selection works. LSF is motivated by the complementary role of intermediate layers of the information they contain about the input. However, this alone does not justify why the consecutive entropy-density drop would be an appropriate signal for selecting an intermediate layer.
  • Additionally, even when we consider LSF purely as a heuristic for intermediate layer selection, I do not think its effectiveness is sufficiently validated. Table 4 compares LSF with alternatives but some of these comparisons do not isolate the contribution of the proposed method. Specifically, if I understand the experimental setup correctly (please feel free to correct me if my reading is wrong), the "Random Layer" baseline uses a randomly selected intermediate representation alone, whereas LSF concatenates the selected intermediate representation with the penultimate representation. This comparison alone cannot conclude whether LSF selects the most suitable intermediate activation. It would also be helpful to clarify if the "Oracle (OOD-informed)" baseline follows the same concatenation procedure. In this regard, I would suggest a comprehensive layerwise evaluation (since the performance of each candidate layer would need to be computed anyway to find the oracle), in which the authors systematically replace the layer selected by LSF with every other candidate intermediate layer. Plotting the resulting FPR95 against the intermediate layer index and highlighting the layer selected by LSF would provide a much clearer picture of the effectiveness of the proposed method.
  • Related to my point about framing, the paper would benefit from an evaluation that focuses more on LSF's influence on different detectors, specifically where it works the best and why that might be the case. I believe the experiment presented in Table 3 is already very relevant, but a more systematic investigation would be better aligned with the main contribution of the paper.
Questions for authors

Questions and minor comments:

  • Why does the approach work better with the Mahalanobis variant?
  • Why would a consecutive entropy drop be relevant to the complementary information contained in intermediate representations?
  • Which model and datasets are used for Figure 1? Are the observations consistent across different architectures and models?
  • The rightmost panel of Figure 2 is not very clear. What do the clusters with different colors represent? The way the paper is presented, I would expect an illustration of the Mahalanobis distance there but I do not think it is clearly conveyed. Teaser figure needs more effort.
  • Figures 4 and 5 are difficult to read, as the distributions overlap and obscure one another.
  • Propositions 2 and 3 feel repetitive and/or trivial and their explanations are unnecessarily lengthy.
  • In Table 4, the penultimate layer only for ConvNext-T outperforms the first row for near OOD setting. Please check the bolding.
  • There are many repetitive and overly defensive phrases throughout the manuscript. The authors repeatedly qualify their claims (especially in sections 3.1-3.4), emphasize what they do not imply or restate things that have already been discussed.
5 comments

We thank the reviewer. The revision refocuses LSF on selection/fusion and strengthens ablations and theory. All references below use the revised manuscript.

W1. Framing and contribution. Section 3.1 selects an intermediate layer using ID-only entropy-density changes (Eqs. 3–5); Section 3.2 concatenates it with the penultimate representation (Eq. 6). Section 3.3 presents Mahalanobis scoring as one instantiation and a tool for analysis, not as a separate proposed detector. Figure 2 summarizes selection, fusion, and scoring. Crucially, Section 4.2, Table 1, pairs penultimate-only and LSF-augmented representations for eight feature-based detectors. With ViT-B/16/ImageNet-1K, mean semantic-OOD FPR95 falls from 38.49% to 30.26% for Maha++ with LSF-B and 42.22% to 33.36% for NNGuide; the gains are not confined to Mahalanobis scoring.

W2. Scientific insight into selection. Section 3.3, Proposition 1(a), Eq. 11, decomposes additional expected OOD–ID squared-Mahalanobis separation into an intermediate-residual mean shift and a covariance-change term:

GF=mr⊤S2−1mr+tr⁡[S2−1(CrQ−CrP)].G_{\mathrm F}=m_r^\top S_2^{-1}m_r+\operatorname{tr}[S_2^{-1}(C_r^Q-C_r^P)].

This fixed-class identity does not assume Gaussianity or independent additive shifts; covariance changes can help or hurt. Under Assumption A1, Proposition 1(b), Eq. 12, bounds the gain using the entropy density of the relevant covariance block. Proposition 1(c), Eq. 13, connects the largest consecutive drop to the greatest one-step decrease in a sufficient-condition reference threshold for equal-width, covariance-matched, block-diagonal population models. The rule does not guarantee the optimal OOD layer. Section 3.3 (Eq. 14) and Appendix A.3 explain why raw intermediate activations may retain additional conditional ID–OOD distinguishability; this need not persist monotonically for pooled, normalized features. Proofs, counterexamples, and practical shift regimes are in Appendices A.2–A.4.

W3. Selection validation. Section 4.4, Table 4, uses the same penultimate-plus-candidate fusion for LSF, Random, and the OOD-informed Oracle. Random averages performance over all candidate layers (uniform selection in expectation); Oracle chooses a candidate using OOD labels. Appendix C.6.1, Table 25, evaluates every ViT-B/16 candidate layer under ImageNet-1K/LT; Table 26 extends the criterion comparison to ImageNet-LT. The ID-only rule selects B4; in the ImageNet-1K sweep, the oracle uses B6 for near/far and B5 for covariate shifts. Table 24 additionally reports selected layers across four backbones and both calibrations.

W4. Broader detector analysis. The controlled paired comparison in Section 4.2, Table 1, shows semantic-OOD FPR95 improvements of 27.21 pp (KPCA), 8.86 pp (NNGuide), and 1.54 pp (KNN). For covariate-shifted ID, Maha++ with LSF-B improves from 70.90% to 53.61% (17.29 pp). Section 3.3 explains conditional sensitivity of intermediate layer fusion; KPCA reconstruction and KNN local geometry offer distinct ways for additional features to affect scores. Appendix C.1, Table 7, gives ImageNet-LT results.

Q1. Why stronger with Mahalanobis? Proposition 1(a–b) (Section 3.3) identifies sensitivity to intermediate residuals along low-ID-variance directions, absent from ordinary Euclidean KNN. Table 1 shows Maha++ with LSF-B has the lowest mean FPR95 for both shift groups, not necessarily every dataset.

Q2. Why the consecutive drop? Section 3.1 (Eq. 4) defines Δl\Delta_l; Proposition 1(c) (Eq. 13) relates it to a one-step threshold reduction under stated assumptions. Appendix A.1, A.2 and A.3 explain the spectral and conditional-information motivations; neither proves that Δl\Delta_l directly measures complementarity.

Q3. Figure 1 and generality. Figure 1 now names ViT-B/16 and ImageNet-1K. Appendix C.4, Figures 6–7, extends CKA and per-layer OOD-separation illustrations across architectures; The last paragraph in Section 4.4 and Table 24, reports selections across calibration sets. The trends vary with backbone and shift type.

Q4. Framework diagram. Revised Figure 2 colors the class clusters and clarifies that Mahalanobis distance is computed in feature space, not from the schematic t-SNE plot.

Q5. Overlapping distributions. Figures 4–5 (Appendix C.3) redraw the ViT-B/16 and ConvNeXt-T score distributions for readability.

Q6. Theory length. Former overlapping propositions are consolidated into Proposition 1(a–c), Section 3.3; proofs and qualifications are in Appendix A.2–A.4.

Q7. ConvNeXt-T bolding. Table 4 correctly recognizes that the penultimate-only score (62.95%) beats LSF-C (63.60%) on near-OOD; the oracle is an OOD-informed reference, not a feasible ID-only selection method.

Q8. Redundancy. Sections 3.1–3.3 and 4.4 are streamlined, with extended derivations moved to Appendix A.

I thank the authors for their substantial effort during the rebuttal period. I appreciate the corrected experimental setup in Table 4 and particularly the comprehensive layerwise evaluation in Table 25. These experiments directly address my concern about the confounded random layer comparison. Additional experiments show that the selected B4 consistently performs among the better candidates.

However, my central concern about the scientific insight provided by the proposed selection criterion is still unresolved.

Additional experiments show that the selection criterion does a good job, but they do not really support why this particular selection is favorable over other selection methods (keeping the selected layer constant at B4 or B5 very consistently improves performance in LSF-C comparisons). This makes me wonder if the entropy drop, similar to that of the selected layer, is high for the well-performing ones. This would change my assessment significantly.

Secondly, I do not think the additional theoretical analysis provides any additional insight into looking for a large entropy drop as a selection mechanism.

I also acknowledge the improvements to the manuscript's framing, figures, and presentation, which address my remaining minor comments.

I believe the authors' rebuttal improved the empirical validation of LSF and I thank the authors for their effort. However, my main concerns still remain. For this reason, I will maintain my original assessment.

We thank the reviewer for recognizing the improvements and for further clarifying the remaining concerns. We address the empirical and theoretical motivations for our layer-selection criterion below.

1. Why is the entropy-density drop a useful selection criterion?

To directly investigate the reviewer's question, we provide a new layerwise analysis reporting entropy density, its change between consecutive layers, and the resulting full covariance (LSF-C) FPR95 for each layer-penultimate pair on ViT-B/16 and CLIP ViT-B/16. Bold mark the best in each row. † marks the layer l⋆l^\star selected by LSF.

B0B1B2B3B4B5B6B7B8B9B10B11Pen.
ViT-B/16
Entropy density ($ \times 10^3$)2.8793.8774.4425.0444.8035.4815.8886.8977.6748.0358.3098.3008.311
Entropy-density drop ($ \times 10^4$)—-10.0-5.6-6.0+2.4 †-6.8-4.1-10.1-7.8-3.6-2.7+0.1-0.1
Near+Far FPR95 (↓\downarrow)37.6535.6733.9232.4330.2828.8328.3130.6433.9540.9253.1964.8738.61*
Covariate FPR95 (↓\downarrow)70.7564.6961.1758.2855.3655.0556.2862.8469.6775.8279.2580.3571.08*
CLIP ViT-B/16
Entropy density (×103\times 10^3)3.0214.1024.4404.6174.8805.1625.2525.6486.3497.3007.8677.9728.177
Entropy-density drop (×104\times 10^4)—-10.8-3.4-1.8-2.6-2.8-0.9 †-4.0-7.0-9.5-5.7-1.1-2.1
Near+Far FPR95 (↓\downarrow)40.5437.8536.6133.3324.6419.3217.6620.1628.5237.2640.9153.2239.00*
Covariate FPR95 (↓\downarrow)68.2564.9159.6254.1348.0746.7844.9345.8550.6560.6269.3578.7668.26*

For ViT-B/16, the selected B4 exhibits the largest entropy-density drop of 2.4×10−42.4\times10^{-4}, whereas B5 and B6 exhibit lower entropy density drops of −6.8×10−4-6.8\times10^{-4} and −4.1×10−4-4.1\times10^{-4}, respectively. Nevertheless, B5 and B6 achieve slightly lower Near+Far FPR95 (28.83% and 28.31%) than B4 (30.28%). For Covariate FPR95, however, B5 achieves a slightly lower value (55.05%) than B4 (55.36%), while B6 performs slightly worse (56.28%). This suggests that a substantial spectral transition can identify favorable ID-only local changes in representation geometry rather than predicting OOD performance of every individual layer.

For CLIP ViT-B/16, the criterion selects B6, which exhibits the largest signed difference (−0.9×10−4-0.9\times10^{-4}), corresponding to the smallest increase in entropy density. B6 achieves the lowest FPR95 in both settings: 17.66% for Near+Far and 44.93% for Covariate. These values improve upon those of both neighboring layers: B5 achieves 19.32% and 46.78%, while B7 achieves 20.16% and 45.85%, respectively.

These results clarify the role of our criterion: it identifies favorable local changes in representation geometry rather than predicting the OOD performance of every individual layer. Unlike selecting the minimum absolute entropy density, which would select B0 in both models, the proposed criterion identifies B4 and B6 using ID-only statistics. OOD data are used exclusively for evaluating these selections, not for determining them.

2. Theoretical motivation for the entropy-density drop.

Proposition 1(a) characterizes the additional expected OOD–ID Mahalanobis separation contributed by an intermediate representation. Proposition 1(b) establishes that, under Assumption A1, lower entropy density strengthens a lower bound on this separation.

Proposition 1(c) provides the direct connection to our selection rule. For equal-width layers under the block-diagonal population-covariance model, a smaller entropy density reduces the minimum OOD variation strength sufficient for positive additional separation at test time.

Concretely, when the threshold decreases, some OOD shifts that do not satisfy the sufficient condition at the preceding layer may satisfy it at the current layer, assuming comparable OOD variation across the two layers. Selecting the largest drop therefore identifies the most favorable local change in this theoretically motivated sensitivity condition. When all drops are negative, the criterion selects the smallest increase in the threshold.

Together, the theoretical analysis explains the relevance of spectral changes to OOD sensitivity, while the new layerwise results illustrate how the ID-only criterion identifies useful intermediate representations without OOD-informed layer tuning.

Reviewer_vZ86OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a post-hoc framework for OOD detection that leverages complementary information from intermediate and penultimate representations of pretrained neural networks. Its main contributions include an entropy density based criterion for selecting an informative intermediate layer using only ID data, a feature concatenation strategy that accommodates representations of different dimensions, and Mahalanobis based OOD scoring with full or block-diagonal covariance estimation. The authors also provide theoretical analysis of the proposed framework regarding cross layer feature decomposition and sensitivity to distribution shifts under specific assumptions.

Critical strengths
  • The paper presents a well-motivated approach to OOD detection by exploiting complementary information from intermediate and penultimate representations of pretrained neural networks.

  • The proposed framework incorporates an entropy density based layer selection strategy that requires only in-distribution data and eliminates the need for auxiliary OOD samples or model retraining.

  • The paper provides theoretical analysis to characterize the proposed feature fusion strategy and its sensitivity to distribution shifts under specific assumptions.

Critical weaknesses
  • The theoretical justification for the proposed entropy density based layer selection criterion is insufficient. The authors can improve by establishing a rigorous connection between entropy density drops and the selection of discriminative features for OOD detection.

  • The theoretical guarantee of non-degradation relies on restrictive assumptions about distribution shifts and does not establish that the proposed method consistently improves practical OOD detection metrics.

  • In Table 1 and Table 2, the values for FPR95 are exactly the same for the baselines RS-Maha and Maha++.

  • The checkpoint "vit_base_patch16_224.augreg2_in21k_ft_in1k" in Table 23 is listed twice, even though it has different paper/table names.

  • In H.2 Public Implementations Used, the "Energy" entries are listed twice.

Questions for authors
  • Could the authors provide a more rigorous theoretical justification for why the largest entropy density drop identifies an intermediate layer containing complementary information beneficial for OOD detection? Is there any theoretical relationship between the proposed selection criterion and OOD detection performance?

  • Could the authors clarify how the assumptions made in Proposition 3 hold in practical OOD scenarios and whether the theoretical results can be extended to performance metrics?

3 comments

We thank the reviewer for the constructive feedback. We revised the theory, clarified the scope of its claims, and corrected the reporting inconsistencies. All references below use the revised manuscript.

W1 / Q1. Why entropy-density drop, and how does it relate to OOD performance? Section 3.1 (Eqs. 3–5) defines spectral entropy density ρl=Hl/Dl\rho_l=H_l/D_l and selects l⋆l^\star by the largest consecutive drop Δl\Delta_l. Section 3.3, Proposition 1(b) (Eq. 12), provides a conditional bound on the additional expected OOD–ID squared-Mahalanobis separation contributed by the intermediate block:

Gv≥Mv+βvgDρv−κv.G_v\ge M_v+\frac{\beta_v g_D}{\rho_v}-\kappa_v.

Here Mv≥0M_v\ge0 is the mean-shift contribution, βv\beta_v lower-bounds OOD variability in the specified non-dominant ID spectral directions (Assumption A1), and κv\kappa_v is the ID contribution. The sufficient condition βv>κvρv/gD\beta_v>\kappa_v\rho_v/g_D connects covariance concentration to potential OOD sensitivity: variation in ID-quiet directions receives greater inverse-covariance weight. Proposition 1(c) (Eq. 13) further shows that, for equal-width layers with matched block-diagonal population covariance, bl−1−bl=(D/gD)Δlb_{l-1}-b_l=(D/g_D)\Delta_l, where bl=Dρl/gDb_l=D\rho_l/g_D is a sufficient-condition reference threshold. Thus the selected transition gives the greatest one-step threshold reduction, not necessarily the best absolute threshold or realized FPR95. The full-covariance bound uses a conditional residual covariance and need not coincide with the marginal spectrum used in selection. Proofs and limitations appear in Appendix A.1–A.2. Empirically, Section 4.4 (Table 4) and Appendix C.6.1 (Tables 25–26) compare the criterion with alternatives and a full candidate-layer sweep; Table 24 reports selected layers across calibrations.

W2 / Q2. Assumptions, practical shifts, and detection metrics. Section 3.3, Proposition 1(a) (Eq. 11), replaces the original unconditional non-degradation argument with a fixed-class expected-contrast decomposition for finite-second-moment distributions. With the intermediate residual r2r_2 conditioned linearly on the penultimate feature (Eq. 9),

GF=mr⊤S2−1mr+tr⁡[S2−1(CrQ−CrP)].G_{\mathrm F}=m_r^\top S_2^{-1}m_r+\operatorname{tr}\left[S_2^{-1}(C_r^Q-C_r^P)\right].

It requires neither Gaussianity nor independent additive OOD shifts. Mean shifts contribute nonnegatively, but residual covariance changes can increase or decrease separation. Appendix A.2 proves the result; Appendix A.3 gives conditional-KL counterexamples, and Appendix A.4 discusses favorable and unfavorable semantic and covariate shifts. This is not a theorem about FPR95/AUROC, which depend on the full score distributions and minimization over classes. Those metrics are evaluated separately in Section 4.3 (FPR95 Tables 1–3), Appendix C.1 (AUROC Tables 8–9) and C.2 (discussion), and the appendix backbone results.

W3. Identical RS-Maha and Maha++ baseline values. Section 4.2, Table 1 and its footnote explain that the ID-only RS-Maha selection gives β=1\beta=1 on the penultimate ViT-B/16 feature, reproducing Maha++ normalization in the matched implementation. For the selected intermediate layer, RS-Maha instead uses β=0.25\beta=0.25, whereas Maha++ uses ℓ2\ell_2 normalization for both blocks; the fused results therefore differ. Appendix D (RS-Maha implementation) specifies the radial transformation and ID-only β\beta search.

W4. Duplicate checkpoint. The checkpoint-to-name mapping in Appendix E, Table 32, now has only one active entry for vit_base_patch16_224.augreg2_in21k_ft_in1k. We deleted the duplicate.

W5. Duplicate Energy implementation. Appendix D lists Energy once in its baseline-method descriptions.

We thank all reviewers for their time and thoughtful feedback!

In response to Reviewer p9sx's latest comments, we have further revised Section 3.3 and Appendix A.1–A.3 to strengthen the theoretical connection between entropy-density drops and potential OOD sensitivity, focusing on spectral sensitivity and complementary information. All references in our responses are to the latest revised manuscript.

As the discussion period draws to a close, we would greatly appreciate any further comments or questions from the reviewers, particularly regarding whether our revisions have addressed their concerns.

Thanks for the authors' response. I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.