acceptodds
Withdrawn from ICLR 2027

Post-Hoc Latent Conflict-Aware Evidence Adjustment for Robust Evidential Deep Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reliable uncertainty quantification is essential for deploying deep learning models in high-stakes settings, where out-of-distribution and adversarial inputs can induce confident but unreliable predictions. Evidential Deep Learning provides efficient uncertainty estimates in a single forward pass, but can still assign high evidential strength to inputs that are poorly supported by the learned representation, such as adversarial inputs. We introduce CLEAR, a lightweight, task-agnostic post-hoc method that improves evidential robustness without retraining or altering the base prediction. Using held-out calibration data, CLEAR characterises the group-conditioned geometry of the model's latent space. At inference, it efficiently generates perturbation views directly in the latent space and measures their conflict relative to the calibrated geometry of the predicted group. High latent conflict indicates unsupported evidence, which CLEAR uses to selectively reduce evidential strength while retaining evidence for latent-consistent inputs. Across classification, regression, and object detection benchmarks, CLEAR substantially improves out-of-distribution and adversarial detection while preserving predictive performance and low computational overhead.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Evidential Deep Learning, Uncertainty Quantification, Post-hoc Uncertainty Estimation, Out-of-Distribution Detection, Adversarial Robustness, Latent Representations

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_k36kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CLEAR, a post-hoc method for improving uncertainty estimation in Evidential Deep Learning (EDL) without retraining or changing the base model's predictions. The authors argue that evidential models may assign strong evidence to OOD or adversarial inputs even when their latent representations are not well supported by the learned distribution.

To address this issue, CLEAR 1) estimates class-conditional latent means and covariances using held-out calibration data, 2) generates multiple perturbed views directly in the latent space, and 3) computes a conflict score from the interactions between the whitened views. When the conflict score is high, CLEAR reduces the evidential strength to increase uncertainty while preserving the original prediction.

The authors evaluate CLEAR on classification, regression, and object detection benchmarks, comparing it against existing evidential and post-hoc uncertainty methods. The results show improvements in OOD and adversarial detection AUROC, particularly on challenging datasets such as ImageNet and CIFAR-10. The authors also report substantially lower inference overhead compared with existing post-hoc methods that require repeated model evaluations.

Critical strengths
  • Efficient post-hoc uncertainty adjustment: CLEAR improves uncertainty estimation without retraining the underlying evidential model or modifying its predictions. Since the perturbations are generated directly in latent space, the method avoids repeated full-network evaluations. This makes the method computationally efficient compared with existing post-hoc approaches such as S-EDL and C-EDL.

  • Strong OOD and adversarial detection performance across multiple tasks: The authors evaluate CLEAR on several classification and regression benchmarks, with additional experiments on object detection. Table 1 shows substantial improvements in adversarial detection AUROC, particularly on CIFAR-10 and ImageNet. These results suggest that the proposed method can provide useful separation between ID and unreliable inputs across different tasks.

Critical weaknesses
  • Unclear justification of the proposed latent conflict mechanism: The authors claim that CLEAR measures the disagreement between perturbed latent views. However, Equation (14) computes the products of squared coordinates across different views, which can produce a non-zero conflict score even when the two views are identical. Moreover, the perturbed views are not passed through the model again, so it is unclear whether the score actually measures local predictive disagreement or simply the distance from the calibrated latent geometry. The paper does not provide a direct comparison with simpler class-conditional distance measures, such as Mahalanobis distance, or a deterministic version of the conflict score without stochastic perturbations. Without these comparisons, it is difficult to determine whether the proposed conflict mechanism provides additional information beyond the underlying latent geometry.

  • Significant degradation in uncertainty calibration: Although CLEAR improves OOD and adversarial detection AUROC, the calibration metrics show substantial degradation. In Appendix C.2, Table 4, CLEAR increases ECE from 5.66% to 33.46% on CIFAR-10 and from 4.56% to 75.74% on ImageNet compared with EDL. The NLL also increases considerably, reaching 6.77 on ImageNet compared with 0.99 for EDL. Since the main objective of the paper is to improve reliable uncertainty estimation, such large calibration errors raise concerns about whether the method produces meaningful uncertainty estimates beyond separating ID and OOD samples. Preserving the predicted class does not necessarily mean that the resulting uncertainty is well calibrated.

Questions for authors
  • Clarification of the latent conflict mechanism: Equation (14) computes interactions between squared coordinates of different latent views, which can be non-zero even when the views are identical. Could the authors clarify how this score measures disagreement between views rather than simply capturing the magnitude of the whitened latent representation? In particular, how would CLEAR compare against a simpler class-conditional Mahalanobis distance or a deterministic version of the proposed score without perturbation sampling? This would help assess the necessity of the proposed conflict mechanism.

  • Clarification of the uncertainty calibration trade-off: Appendix C.2 reports substantial increases in ECE and NLL, particularly on CIFAR-10 and ImageNet. Could the authors clarify why these degradations are acceptable for reliable uncertainty estimation? Also, is it possible to retain the improvements in OOD and adversarial detection while maintaining reasonable calibration performance? This would help assess the practical usefulness of CLEAR.

  • Clarification of adversarial robustness: The paper evaluates CLEAR under several adversarial attacks, including PGD and AutoAttack. However, it is unclear whether these attacks explicitly target the adjusted uncertainty produced by CLEAR. Could the authors clarify whether an adaptive attacker minimizing the latent conflict score was considered? This would help assess whether the reported improvements persist against attacks targeting the uncertainty estimation mechanism itself.

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

This paper proposes CLEAR, a post-hoc method for improving the robustness of Evidential Deep Learning (EDL) against out-of-distribution and adversarial inputs. CLEAR estimates class-conditional latent geometry using calibration data and measures latent conflict through multiple perturbed views. Based on the resulting conflict score, it selectively reduces evidential strength without retraining the model or changing its predictions. Experiments across classification, regression, and object detection demonstrate improved OOD and adversarial detection with relatively low computational overhead.

Critical strengths

(1. Practical and computationally efficient post-hoc framework.) CLEAR improves uncertainty-based OOD and adversarial detection without retraining the underlying model or modifying its point predictions. It introduces relatively low inference overhead compared to existing post-hoc methods. This makes the proposed approach practically appealing, particularly for applications where retraining or repeated inference is computationally expensive.

(2. Broad empirical evaluation with promising results.) The paper provides extensive experiments across classification, regression, and object detection, considering diverse OOD datasets and adversarial attacks. CLEAR demonstrates improvements in OOD and adversarial detection across many evaluated settings, particularly in adversarial robustness.

Critical weaknesses

(1. Insufficient justification of the latent conflict mechanism and the necessity of multiple views.)

While I think the central hypothesis of CLEAR is reasonable—that reliable evidence should be supported by the calibrated class-conditional latent geometry—I find several key methodological choices insufficiently motivated.

First, it is unclear why the proposed pairwise conflict score in Eq. (14) should be interpreted as latent conflict. The score is defined through cross-dimensional products of squared whitened latent coordinates, but I find it difficult to understand intuitively how these interactions represent inconsistency with the calibrated latent geometry. In particular, the score can be nonzero even when two latent views are identical, suggesting that it does not directly measure disagreement between views. Although the authors interpret large cross-dimensional interactions as latent conflict, a clearer explanation of why these interactions indicate unsupported evidence would be necessary to justify this formulation.

Second, the necessity of generating multiple stochastic latent views is unclear. Under the authors' central hypothesis, the compatibility of a test sample with the calibrated latent geometry could seemingly be assessed directly from its original representation (h(x)), for example, using the Mahalanobis distance. Multiple views are commonly motivated in uncertainty estimation by measuring predictive variability or disagreement across perturbed inputs. However, CLEAR does not pass the generated latent views through the evidential head or examine the dispersion of their predictions or evidence. Instead, it computes pairwise interactions between perturbed latent representations. Consequently, I find it difficult to understand what additional information these multiple views provide beyond the original latent representation and the calibrated geometry, or why they are necessary for the proposed uncertainty adjustment.

Overall, although the empirical results are promising, the paper does not sufficiently explain how these design choices follow from its central hypothesis. I believe a more convincing conceptual justification of the proposed conflict measure and the role of multiple latent views is necessary to establish the methodological contribution of CLEAR.

(2. Missing comparisons with closely related and simpler baselines.) The paper compares CLEAR against several EDL-based methods but omits DAEDL [1], which is closely related to the proposed approach in leveraging latent-space density information to improve evidential uncertainty estimation.

Additionally, since CLEAR already estimates class-conditional latent means and covariances, a simple Mahalanobis-distance-based evidence adjustment would be a particularly relevant baseline. Without these comparisons, it remains unclear whether the reported improvements arise from the proposed latent conflict mechanism or simply from incorporating information about the latent distribution of ID samples. These missing comparisons make it difficult to assess the necessity and contribution of the proposed mechanism over existing or simpler alternatives.

References [1] Uncertainty estimation by density aware evidential deep learning, ICML 2024

Questions for authors

(1. Interpretation of the latent conflict score.) Could the authors provide a more intuitive or theoretical justification for why the cross-dimensional interactions in Eq. (14) should be interpreted as latent conflict? In particular, since the score can be nonzero even when two latent views are identical, it is unclear how this quantity reflects disagreement between views or incompatibility with the calibrated latent geometry. Clarifying the relationship between the proposed score and unsupported evidence would help establish the validity of the central mechanism.

(2. Necessity of multiple latent views.) What additional information do multiple stochastic latent views provide beyond the original latent representation (h(x)) and the calibrated class-conditional geometry? Since CLEAR does not evaluate predictive disagreement across the generated views, the motivation for multiple views remains unclear. Furthermore, the expected conflict score can be computed analytically without stochastic sampling. Could the authors clarify the necessity of this design and, if possible, provide an ablation comparing the proposed multi-view mechanism with a single-view or analytically computed conflict score?

(3. Comparison with related and simpler alternatives.) Could the authors provide a comparison with DAEDL [1], which also leverages latent-space density information for evidential uncertainty estimation? Additionally, since CLEAR already estimates class-conditional means and covariances, a Mahalanobis-distance-based evidence adjustment appears to be a natural and computationally inexpensive alternative. Comparing these approaches under the same evaluation protocol would help determine whether the proposed latent conflict mechanism offers benefits beyond incorporating latent distribution information.

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

This paper proposes a novel approach to adjusting the evidence produced by evidential learning frameworks for both classification and regression tasks. The proposed method introduces an evidence scaling factor that modifies the evidence and, consequently, the Dirichlet distribution parameters and predicted class probabilities in classification tasks, as well as the evidence-related parameters of the Normal-Inverse-Gamma (NIG) distribution in regression tasks. The evidence scaling factor is derived from a conflict score, which quantifies disagreement in the latent space across perturbed views of latent representations generated by a pretrained evidential model. For in-distribution (ID) samples, the conflict score is expected to approach zero, resulting in an evidence scaling factor close to unity and, therefore, minimal adjustment to the original evidential outputs. In contrast, for out-of-distribution (OOD) samples, the conflict score determines the extent of evidence adjustment based on the disagreement observed in the latent space. The proposed method is evaluated across seven ID/OOD dataset combinations for classification and five ID/OOD dataset combinations for regression. The paper further includes an evaluation on object detection tasks and extensive ablation studies to support the effectiveness of the proposed approach.

Critical strengths

The paper has several notable strengths, and I would like to highlight a few that I found particularly compelling.

  1. Well-written and clearly motivated: The paper is very well-written, providing a clear discussion of the background and motivation. I particularly appreciated the inclusion of the fundamental equations of evidential deep learning for both classification and regression tasks. The authors' decision to prioritize this theoretical background in the main paper, rather than presenting extensive ablation studies, seems appropriate, as it provides readers with the necessary foundation to understand the proposed method.

  2. Novel and intuitive methodology: The proposed method introduces a novel approach to modulating evidence based on the stability of latent representations. Rather than measuring disagreement in the output probability space, the method quantifies disagreement in the latent space across perturbed views of the learned representations. The underlying intuition is compelling: in-distribution (ID) samples are expected to exhibit relatively low variation under perturbations because they are well represented in the training distribution, whereas out-of-distribution (OOD) samples are likely to exhibit greater variation because their representations may not have been learned robustly. The separation between ID and OOD samples is clearly demonstrated through robustness experiments, and the experimental design, along with the carefully chosen implementation strategies, effectively captures this distinction.

  3. Extensive experimental evaluation: The experimental evaluation is extensive, if not exhaustive, covering classification, regression, and object detection tasks. I particularly appreciated the inclusion of object detection experiments and the thoughtful experimental design involving the EMM baseline, which helps isolate and demonstrate the contribution of CLEAR to uncertainty estimation. Furthermore, evaluating the method under different perturbation settings is an appropriate choice, as these scenarios directly examine the robustness of the learned representations and the effectiveness of uncertainty-aware evidence adjustment. Collectively, these experiments provide strong empirical support for the proposed method.

Critical weaknesses

None

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.