acceptodds
Under review as a conference paper at ICLR 2027

Beyond Predictive Criteria: Optimization and Downstream Uncertainty

Abstract

Optimization choices are typically guided by predictive criteria such as accuracy, generalization, and calibration, yet many downstream uncertainty estimators operate on learned representations. We ask how uncertainty readouts respond to representations produced by predictively favorable optimization changes. Using the perturbation radius of Sharpness-Aware Minimization (SAM) as a controlled training path, we find a selective divergence: accuracy and calibration improve, output-based OOD scores remain stable or improve, and kNN improves, whereas covariance-sensitive Mahalanobis scoring and Deep Deterministic Uncertainty (DDU) can deteriorate at large SAM radii. We investigate this degradation using Mahalanobis scoring as a tractable case. Within each checkpoint, we vary the strength of covariance weighting while holding the learned representation, class centroids, and covariance eigensystem fixed. Under large-radius SAM, stronger inverse-covariance weighting can become substantially detrimental to OOD detection. Directional analysis links this behavior to a covariance–contrast mismatch in which later low-variance eigendirections receive increasing emphasis despite providing weaker ID–OOD contrast. Crucially, an unweighted readout over the same representation retains strong OOD discrimination as these directions are added, showing that the degradation does not require a uniform loss of OOD-discriminative structure. This covariance-sensitive degradation under large-radius SAM is not universal across sharpness-aware optimizers or ID training distributions, while the broader separation between predictive improvement and representation-dependent uncertainty also appears under SWA and in a ViT/ImageNet-200 setting. Together, these results show that uncertainty readouts can respond differently across representations produced under different optimization conditions and, in the Mahalanobis case, that detector degradation can arise from detrimental inverse-covariance weighting even when useful OOD-discriminative structure remains accessible.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Optimization, Uncertainty Estimation, Out-of-Distribution Dection, Representation Geometry, Mahalanobis Distance

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_2GypOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces an explainability-based framework for assessing the reliability and effectiveness of uncertainty estimates produced by predictive models.

Critical strengths
  1. It provides a novel perspective on understanding the effectiveness of uncertainty estimation.
Critical weaknesses
  1. Lack of clarity and readability. The paper is difficult to follow, particularly in its motivation, methodological analysis, and interpretation of the experimental results. The presentation should be more concise and logically structured to make the central arguments easier to understand.

  2. Insufficient justification of the problem's significance. The practical relevance and scientific significance of the research question are not sufficiently established. It remains unclear why the observed divergence between predictive performance and downstream uncertainty estimation constitutes an important problem beyond the specific experimental settings considered.

  3. Limited completeness and depth of the study. The current investigation appears insufficiently developed for publication. Although the paper provides empirical observations and diagnostic analyses, it lacks a sufficiently general theoretical explanation or actionable methodological insights to establish a substantial contribution.

  4. Unclear conclusions and key takeaways. The main conclusions of the paper are not sufficiently clear. While the authors demonstrate that different uncertainty estimators respond differently to optimization-induced representation changes, it remains unclear what generalizable insights can be drawn from these findings and how they can guide the design, selection, or evaluation of uncertainty estimation methods.

Questions for authors
  1. Is this framework suitable for other uncertainty models, such as Bayesian models?
1 comment

We thank the reviewer for the constructive feedback. We have revised the paper to clarify its research question, main finding, and practical significance.

1. Clarity and research focus. Our study is not an explainability framework or a proposal for improving SAM. We investigate how the discriminative effect of covariance weighting on OOD detection changes across training conditions. SAM provides a controlled training axis, while Mahalanobis distance allows us to isolate covariance weighting within each trained checkpoint.

2. Scientific significance. Section 4.1 identifies a training-dependent difference that is not apparent from comparing final Mahalanobis scores alone. On CIFAR-100, equal weighting favors large-radius SAM over SGD (0.9237 vs. 0.9098 AUROC), whereas inverse-covariance weighting reverses their relative performance (0.8787 vs. 0.9076). Thus, the same weighting intervention has substantially different discriminative effects across training conditions.

3. Depth of analysis. Section 4.2 examines how weighted score contributions accumulate. Under large-radius SAM, later inverse-weighted directions increase mean ID–OOD separation but increase variability faster as AUROC declines. A Gaussian reference calculation illustrates how this imbalance can arise. Section 4.3 then shows that mean separation and variability alone do not determine AUROC: after L2 normalization and covariance refitting, normalized SAM achieves higher AUROC than SGD despite lower standardized separation. We analyze the corresponding score-difference distributions and pairwise rankings.

4. Conclusions and practical implications. Experiments across training procedures, covariance regularization, OOD datasets, and ViT representations show that inverse weighting can be beneficial or detrimental depending on the setting. The practical implication is to evaluate covariance weighting together with the trained representation, rather than assuming that a scoring rule effective for one training condition will remain effective for another. When representative OOD evaluation data are available, comparing equal and inverse-covariance weighting within the same checkpoint provides a diagnostic of weighting-induced gains or losses.

Regarding Bayesian models, our intervention concerns feature-based detectors with fitted covariance statistics and does not directly extend to Bayesian posterior uncertainty. Investigating an analogous interaction in Bayesian uncertainty estimation would require a separate formulation and validation.

Reviewer_DBUqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines empirically how uncertainty readouts change in response to changing the perturbation radius ρ\rho of Sharpness-Aware Minimization (SAM). They show how different metrics behave differently via numerous experimental results: for example accuracy and calibration improve, whereas covariance-sensitive Mahalanobis scoring and Deep deterministic uncertainty (DDU) can degrade as ρ\rho increases.

Critical strengths

The authors provide a variety of experimental results, showing the "selective divergence" which occurs in SAM.

Critical weaknesses

For a paper for which the main results are using SAM, there seems to be glaring omissions of details of the method. While this might be common knowledge, it is probably necessary to properly introduce SAM, defining all related terminology such as the perturbation radius, which is never properly defined in the main text. I also could not find it in the appendix.

The results of this paper look to be purely empirical, where the authors show what they term a "selective divergence". While the results themselves are interesting, I am unsure about their significance. It doesn't seem to suggest a principled way to choose ρ\rho, or a theoretical explanation for the selective divergence.

Questions for authors

My question is regarding the significance of the experimental results. The main contribution of the paper looks to be the experimental results, but I am unable to appreciate its importance in relation to SAM. Does the "selective divergence" give us insight in choosing ρ\rho such that degradation does not happen? Even in a simple toy example, are we able to determine when the degradation would occur to give some theoretical insight to this divergence phenomenon?

1 comment

We thank the reviewer. We have refocused the paper on how the discriminative effect of covariance weighting changes across training conditions, rather than treating detector divergence under SAM as the main contribution.

1. SAM's role. Section 3.1 now defines the SAM objective and perturbation radius ρ\rho. We vary ρ\rho while holding the architecture, data, and remaining training protocol fixed. SAM serves as a controlled training axis, not as a proposed OOD detection method.

2. Main finding. In Section 4.1, we fix the features, class centroids, covariance eigensystem, and active subspace within each checkpoint, varying only covariance weighting. On CIFAR-100, equal weighting favors large-radius SAM over SGD (0.9237 vs. 0.9098 AUROC), whereas inverse-covariance weighting reverses the ordering (0.8787 vs. 0.9076). The corresponding weighting penalties are 0.0449 for SAM and 0.0021 for SGD, demonstrating a substantial training-dependent difference in the effect of the same weighting rule.

3. Theoretical insight and empirical explanation. Section 4.2 provides a Gaussian reference calculation. A direction with no OOD mean shift adds no mean score separation but contributes variance under inverse-covariance weighting. Consequently, adding such directions can reduce standardized separation even with known covariance. Consistent with this mechanism, under large-radius SAM, later inverse-weighted directions increase empirical mean separation but increase variability faster as AUROC declines. Section 4.3 further shows that mean and variance alone cannot determine the final rankings: after L2 normalization and covariance refitting, the CIFAR-100 weighting penalty falls from 0.0449 to 0.0029, while normalized SAM achieves higher AUROC than SGD despite lower standardized separation.

4. Choosing ρ\rho. These results do not yield a universal safe radius. Weighting effects vary across training procedures, feature preprocessing, and OOD distributions, including ViT settings where inverse-covariance weighting improves discrimination. Instead, when representative OOD evaluation data are available, the fixed-checkpoint comparison of equal and inverse-covariance weighting provides a direct diagnostic of weighting-induced gains or losses. This is the practical implication of our analysis, rather than a universal rule for selecting ρ\rho.

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

The authors study how training affects OOD detection by comparing SGD with SAM, which trains a network to keep its loss low under small weight perturbations. In the main experiments, SAM with larger perturbation radii improves accuracy and calibration compared with SGD. This in turn makes Mahalanobis and DDU worse at OOD detection. To investigate this drop, the authors keep the features fixed while changing only the weights in the distance formula. They find that equal weights over the same retained covariance eigendirections can outperform Mahalanobis weighting, which emphasizes directions with little ID variation around class centers.

Critical strengths
  • The test is clean with respect to the related factors : within each checkpoint, the authors keep the features, class centers, and covariance eigensystem fixed while varying only the distance weights, isolating their effect on OOD AUROC

  • Good ID/OOD separation with equal weights shows that useful information remains in the same features, even when Mahalanobis detection degrades

  • The comparative study is conducted fairly; the appendix checks other training methods, feature normalization, and batch sizes, and reports cases where Mahalanobis does not get worse

Critical weaknesses
  • Figure 3(b) shows that, under SAM with large perturbation radii, Mahalanobis gets worse as directions with smaller within-class variance are added. Since the distance to each class center sums contributions across directions, a key issue is how the mean ID/OOD score gap compares with score variability as directions are added. A simple Gaussian model could characterize this balance using concentration results in high dimensions (" High-Dimensional Probability: An Introduction with Applications in Data Science", Vershynin, 2026 ), distinguishing effects present with a known covariance from those due to estimation

  • In the main experiments, SAM with large perturbation radii makes Mahalanobis worse at OOD detection but improves kNN compared with SGD. The kNN score uses the distance to the k-th nearest ID training feature, after scaling all feature vectors to length one. Comparing these distance distributions for ID and OOD test inputs would help explain why kNN improves under SAM. The reported measures of feature spread describe each class as a whole, rather than these local distances. This analysis could also clarify the roles of dimension, k, and the number of ID training examples

  • The paper does not show whether Equation (6) also explains the effect of feature normalization. In Appendix F.1, the authors scale both SAM and SGD features to length one and refit the class centers and the covariances, which makes the Mahalanobis drop at large SAM radii disappear relative to normalized SGD. If one recomputes the directional contrasts and the prevalence of the condition in Equation (6) before and after normalization, that would help understand whether the same explanation covers the drop and also its disappearance.

  • Krumpl et al. show that OOD performance depends on both training and detector choice, and Li et al. show that changing the covariance in the score can improve OOD detection without changing the features. The authors do discuss both papers in Sections 2.3 and 2.2, but a more direct comparison would clarify what explanatory insight is added by studying covariance weighting across training conditions.

Questions for authors

see above as well as

  • Could a simple Gaussian model show that adding directions with little ID/OOD separation hurts Mahalanobis detection? How must OOD shifts in mean or covariance scale with dimension to maintain separation between the ID and OOD score distributions?

  • Sun et al. (2022) give in "Out-of-Distribution Detection with Deep Nearest Neighbors" an explanation based on feature lengths (Section 5, Figure 4). Can the authors formalize the conditions under which scaling features to length one helps Deep kNN OOD detection?

2 comments

We sincerely thank the reviewer for the careful and constructive review. Your suggestions prompted new analyses that helped us understand our results more clearly and substantially strengthen the revised paper.

W1. Separation and variability. Following your suggestion, we analyzed how mean ID–OOD separation and score variability accumulate as covariance eigendirections are added. Under large-radius SAM, later inverse-weighted directions increase mean separation, but variability grows faster and AUROC declines (Figure 3). At k=112k=112, the mean gap is 99.21, combined SD is 92.86, and AUROC is 0.8797. Our Gaussian calculation shows that this imbalance can arise with known covariance: a zero-shift direction adds no mean separation but contributes variance 44 under inverse weighting, versus 4λi24\lambda_i^2 under equal weighting. The empirical weighting intervention holds the fitted covariance fixed within each checkpoint. These analyses are now central to Section 4.2.

W2. Why kNN improves. Your suggestion to compare ID and OOD nearest-neighbor distance distributions revealed why kNN improves under SAM. The lower tail of OOD distances shifts farther from ID training features relative to the ID acceptance threshold, reducing the fraction of OOD inputs incorrectly accepted as ID. On CIFAR-100, the OOD fifth percentile increases from 0.128 to 0.240, while the threshold changes only from 0.397 to 0.432. Consequently, FPR95 decreases from 46.4% to 35.2%, and AUROC increases from 0.9160 to 0.9346. This pattern holds across all four OOD datasets and three seeds. We also checked the sensitivity to kk (10, 25, 50, 100), ID reference-set size (1–100%), and PCA dimension (8–640); SAM retained a higher normalized-kNN AUROC than SGD in every setting (CIFAR-100: +0.016 to +0.019). Participation ratios of the normalized training features were similar (8.69 for SGD and 8.90 for SAM), suggesting that effective dimensionality alone does not explain the improvement. We incorporated the distance-distribution finding into Section 3.3.

W3. After normalization. We recomputed the directional mismatch diagnostic after L2 normalization and covariance refitting. Its prevalence decreases from 57.51% to 30.23% for SGD and from 85.08% to 57.06% for SAM. However, SAM still has higher mismatch prevalence while achieving higher AUROC (0.9230 vs. 0.9138). Thus, the mismatch condition alone does not explain the recovery. We further found that normalized SAM has lower standardized separation (1.1884 vs. 1.2705) despite higher AUROC. This led us to analyze ID–OOD score-difference distributions and pairwise rankings in Section 4.3.

W4. Prior work. We removed an inaccurate description of Li et al. (2025) from the original manuscript: their work concerns the mismatch between supervised classification and OOD detection objectives, rather than covariance-weighting interventions. Krumpl et al. examined training-dependent detector behavior, while Janiak et al. studied representation geometry, directional separation, and covariance-related score variability. Our contribution is the fixed-checkpoint demonstration that the discriminative effect of the same covariance-weighting rule differs across training conditions, reversing the relative OOD performance of SGD and SAM. We further examine this difference through cumulative score separation, variability, and ID–OOD pairwise rankings.

Q1. Gaussian scaling. For ID N(0,Ik)\mathcal{N}(0,I_k) and OOD N(m,R)\mathcal{N}(m,R), the squared-norm score gives G=tr⁡(R)−k+∥m∥2G=\operatorname{tr}(R)-k+\lVert m\rVert^2 and V=2k+2tr⁡(R2)+4m⊤RmV=2k+2\operatorname{tr}(R^2)+4m^\top Rm. Constant-order standardized separation requires ∥m∥2=Θ(k)\lVert m\rVert^2=\Theta(\sqrt{k}) for a pure mean shift or ϵ=Θ(k−1/2)\epsilon=\Theta(k^{-1/2}) for an isotropic covariance shift R=(1+ϵ)IkR=(1+\epsilon)I_k.

Q2. When normalization helps kNN. Building on the explanation of Sun et al., we decompose the raw distance as ∥z−x∥2=(rz−rx)2+2rzrx(1−u⊤v)\lVert z-x\rVert^2=(r_z-r_x)^2+2r_zr_x(1-u^\top v), where z=rzuz=r_zu and x=rxvx=r_xv have unit directions u,vu,v. Normalization removes the radial terms and improves AUROC when it corrects more ID–OOD pairwise rankings than it reverses. Under SAM, OOD features have systematically shorter norms than ID features (median OOD/ID ratios 0.62–0.79, vs. 0.86–0.88 under SGD). On CIFAR-100, normalization corrects 1.78% of pairs misranked by raw kNN while reversing 0.83%, yielding an AUROC gain of 0.0095; under SGD, the two fractions nearly cancel (0.63% vs. 0.49%). This provides empirical support for the normalization effect described by Sun et al.

We are especially grateful that your questions led to substantive improvements in our analysis and helped us present its central findings more clearly.

Thank you for your response and effort. My overall assessment is that this is a useful empirical study, but the explanations do not yet cover comprehensively enough both the cases where Mahalanobis gets worse and those where it does not. A clearer analysis of these different outcomes would strengthen the paper. My confidence in the study went up following the rebuttal, but I think it is still rash for publication in its current form.

Reviewer_53y3OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper investigates the accuracy and calibration behavior of the Sharpness-Aware Minimization (SAM) baseline under various perturbation settings.

Critical strengths

The paper presents a comprehensive analysis across multiple datasets under various perturbation settings.

Critical weaknesses

The major weaknesses of the paper are as follows:

  1. Lack of clarity and insufficient background: The paper is not well-written and lacks a clear problem formulation. Furthermore, it does not provide sufficient background on Sharpness-Aware Minimization (SAM), which serves as the central focus of the entire analysis. Given that the paper investigates the behavior of SAM under different perturbation settings, it is essential to introduce the fundamental concepts and formulation of SAM. Including this background would make the paper more self-contained and accessible to readers unfamiliar with the method.

  2. Lack of methodological novelty: The paper does not introduce any novel methodological component. Instead, it primarily analyzes the performance of an already established method, SAM, across different datasets and perturbation settings. While an empirical analysis can constitute a meaningful research contribution, the paper does not clearly establish what new insights or understanding emerge from these experiments beyond the existing literature.

  3. Lack of coherent findings and conclusions: Despite presenting extensive experimental results, the paper fails to establish a coherent narrative or derive consistent insights from the analysis. For instance, Figure 1 demonstrates substantially different performance patterns across datasets, making it difficult to identify a generalizable trend or a clear takeaway. Rather than systematically interpreting these differences and explaining their underlying causes, the paper largely presents individual observations about the experimental curves without connecting them to a broader hypothesis or conclusion. Consequently, the analysis reads more like a collection of disconnected, descriptive interpretations of experimental results than a cohesive scientific investigation. The authors should synthesize their findings, identify meaningful patterns or exceptions, and provide a clearer explanation of what these observations reveal about the accuracy and calibration behavior of SAM.

Questions for authors

None

1 comment

We thank the reviewer for highlighting the need for a clearer research question, sufficient background, and more coherent conclusions. We have revised the Introduction and Sections 3–5 accordingly, and Section 3.1 now explicitly introduces the SAM objective and perturbation radius ρ .

We would particularly like to clarify the focus of our study. SAM serves as a controlled training axis rather than the primary subject of the analysis. Our central question is how the effect of covariance weighting on Mahalanobis OOD discrimination changes across training conditions.

To investigate this, Section 4.1 fixes the features, class centroids, covariance eigensystem, and active subspace within each checkpoint, varying only covariance weighting. On CIFAR-100, equal weighting favors large-radius SAM over SGD (0.9237 vs. 0.9098 AUROC), whereas inverse-covariance weighting reverses their relative performance (0.8787 vs. 0.9076). This is our main empirical finding, rather than the initial observation of divergent detector behavior under SAM.

Sections 4.2–4.3 further examine this difference through cumulative score separation, variability, and ID–OOD pairwise rankings. Section 5 establishes its dependence on training and evaluation settings. Together, these analyses show that the same covariance-weighting rule can have substantially different discriminative effects across training conditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.