acceptodds
Under review as a conference paper at ICLR 2027

Lorentz Evidential Deep learning for Out-of-distribution Detection

Abstract

Evidential Deep Learning (EDL) provides an efficient paradigm for open‑set recognition by quantifying uncertainty in a single forward pass. However, existing Euclidean‑based EDL methods struggle to directly adapt to hyperbolic representation learning, lacking standardized and interpretable rules for evidence construction and valid correlations between Dirichlet parameters and manifold geometric confidence, which leads to inherent limitations in hyperbolic evidence modeling. To address these issues, we present the Lorentz Evidential Deep Learning (LEDL) framework to refine evidence modeling in hyperbolic space. Specifically, we first formulate the Lorentz‑decoupling constraint to screen valid Lorentz embedding components from the perspective of semantic reliability and provide geometric rationale for reliable hyperbolic evidence construction. Second, we build a dedicated hyperbolic evidence inference pipeline based on the proposed constraint to achieve stable and standardized Dirichlet parameter generation. Finally, we develop Lorentz Prototype‑Guided Evidence Alignment to establish explicit mappings between evidence and manifold geometric confidence, enabling close alignment between evidential modeling and hyperbolic geometric properties. Extensive experiments demonstrate that the proposed LEDL achieves competitive open‑set detection performance and robustness compared to existing baseline methods.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Evidential Deep Learning, Open-set Recognition, Out-of-distribution Detection

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_KQ1MOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework that integrates hyperbolic representation learning with edl for improving OOD detection and uncertainty estimation. The authors explain that conventional EDL methods arenot suited to hyperbolic representations. The paper proposes three contributions. It introduces a constraint that constructs dirichlet evidence from the spatial components for avoiding geometric interference. It also develops a hyperbolic evidence inference that maps these representations into valid dirichlet concentration parameters.

Critical strengths

The paper shows competitive OOD detection performance against multiple benchmarks, with improvements on CIFAR-100 and long-tailed datasets. These improvements are supported by experiments across multiple random seeds and complementary metrics and evaluation metrics.

It also shows an interesting problem: how to construct dirichlet evidence from representations while consdiering their geometric structure. This proposed parameterization and LPEA illustrate an integrated architectural approach to connecting geometric representations with evidential uncertainty.

Critical weaknesses

The central argument and claim of the paper is based on the separation between radial information and class-discriminative semantics rather than a rigorous theoretical demonstration. Additionally the Lorentz time coordinate is deterministically related to the norm of the spatial components, it means excluding it does not eliminate the radial information from the evidence network. Then the proposed constraint does not establish the contribution of the semantic separation.

Also the main goal of the LPEA framework is to align evidential confidence with manifold geometry. But its loss only penalizes ground-truth evidence under a threshold, without enforcing a relationship between geometric distance and evidence strength. Furthermore, the experimental evaluation focuses on OOD detection metrics only which shows discrimination but do not establish the claimed of geometry–uncertainty alignment or probabilistic calibration.

Questions for authors

On of the paper contributions is that excluding the time radial coordinate prevents semantic contamination in evidence construction. However the time coordinate is completely related to the norm of the spatial components. How does the proposed spatial only parameterization guarantee that radial information does not influence the resulting of dirichlet evidence?

To what extent are the observed OOD detection improvements attributable to Lorentz representations rather than the additional evidence-alignment regularization? Could the authors provide a controlled comparison between Euclidean and Lorentz evidential models using the same backbone, evidence network, prototype mechanism, and training objective?

Table 3 of the paper shows reduced in distribution classification accuracy againts to baseline methods (under class imbalance). Why does the proposed method improve OOD detection but reduce classification accuracy? Whether OOD detection can be improved without reducing classification accuracy?

Are the improvements in OOD detection mainly because of using lorentz representations or the proposed evidence alignment method? Could the authors compare euclidean and lorentz models under the same experimental settings to show the contribution of each one?

5 comments
AuthorsOpenReview ↗

Responses to Reviewer KQ1M(Part 1)

To facilitate the reviewer in identifying which weakness or question is being addressed, we have organized the index numbers as follows:

Critical Weaknesses: Weakness 1: The central argument ...

Weakness 2: Also the main goal of the LPEA ...

Questions For Authors:

Question 1: On of the paper contributions is ...

Question 2: To what extent are the ...

Question 3: Table 3 of the paper shows ....

Question 4: Are the improvements ....

Response to Weakness 1 & Question 1

We appreciate the reviewer raising this precise point. That is entirely correct: given the hyperboloid constraint z0=∥zs∥22+kz_0 = \sqrt{\|z_s\|_2^2 + k}, the time coordinate z0z_0 and the spatial norm are mathematically tied. Our method does not "erase" radial information in an absolute, information-theoretic sense; rather, the core rationale is rooted in optimization dynamics and numerical stability.

Specifically, because z0z_0 is strictly bounded below by k>0\sqrt{k} > 0, it injects a permanent, non-zero scalar offset into the feature space. When the full [z0,zs][z_0, z_s] vector is fed directly into the evidence network—as tested in our Strategy-1 ablation in Table 4—this large positive scalar tends to dominate both the representations and the backpropagation gradients. Under skewed or heavy-tailed distributions like CIFAR100-LT, this common-mode offset overwhelms the fine-grained class semantics in zsz_s, directly triggering the optimization pathologies that cause the sharp performance drop.

Restricting parameterization strictly to zsz_s bypasses this numerical bias and gradient dominance while preserving the local manifold geometry needed for evidence estimation. We have updated Appendix B.1 in the revised manuscript to clarify this optimization-focused perspective (see the blue-highlighted text in Appendix B.1).

AuthorsOpenReview ↗

Responses to Reviewer KQ1M(Part 2)

Response to Weakness 2

We appreciate the reviewer’s valuable suggestions regarding alignment. We further explain the design rationale of Lali\mathcal{L}^{ali} and discuss why direct distance-to-evidence regression brings unique optimization difficulties in hyperbolic spaces, while our threshold-based formulation enables sound geometric alignment.

1. Considerations on Direct Distance-to-Evidence Regression in Hyperbolic Space: One simple scheme is to define a fixed functional mapping, e.g., letting Dirichlet evidence scale inversely with Euclidean or pseudo-Riemannian prototype distance. However, the Lorentz model of hyperbolic geometry has nonconstant negative curvature, and spatial volume expands exponentially rather than linearly. A rigid distance-evidence penalty applied across the entire manifold may trigger severe optimization conflicts:

  • Metric Distortion: Distances in hyperbolic space are strongly distorted near manifold boundaries. A uniform penalty encourages the evidence network to follow flat Euclidean metric rules, weakening the hierarchical representational capacity that motivates adopting the Lorentz manifold.
  • Intra-Class Dispersion vs. Inter-Class Separation: For complex or long-tailed data, discriminative samples of the same class do not converge to a single point; they form clusters with natural geometric spread. A strict distance-evidence constraint penalizes this reasonable intra-class variation, driving the model to overcompress features and impair closed-set decision boundaries.

2. How Lali\mathcal{L}^{ali} Achieves Robust Geometric-Evidential Alignment: Rather than enforcing an overly rigid global regression constraint, Lali\mathcal{L}^{ali} adopts a geometry-aware strategy based on large-deviation subsets (Ictop\mathcal{I}_c^{top}).

  • Asymmetric Soft-Bounding: The threshold tct_c acts as a geometric boundary condition. It does not require a fixed proximity for valid evidence, but imposes a minimum acceptable evidence level for in-distribution samples that sit within reasonable geometric proximity to their class prototypes. By penalizing only samples that fall below this threshold, Lali\mathcal{L}^{ali} fixes abnormal evidence collapse and retains the continuous distance–evidence relationship over the rest of the manifold. This yields stable alignment compatible with the underlying topology.

Table 9: Ablation study for loss Lali\mathcal{L}^{ali} using CIFAR100 and CIFAR100‑LT (η=0.01\eta=0.01) as ID datasets. (Note: Due to web response length limits, C100 denotes CIFAR-100, and C100LT denotes CIFAR-100-LT (η=0.01\eta=0.01).)

MethodC100C100C100C100C100LTC100LTC100LTC100LT
OODOODOODIDOODOODOODID
FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$
Ours(w/o Lali\mathcal{L}^{ali})40.29 ± 1.1487.88 ± 0.5291.52 ± 0.1880.13 ± 0.1756.07 ± 1.7876.05 ± 1.0484.92 ± 0.7250.89 ± 0.46
Ours36.50 ± 2.0488.32 ± 0.6992.24 ± 0.4380.10 ± 0.1348.25 ± 3.3379.08 ± 1.0388.36 ± 0.9049.88 ± 0.62

3. Addressing the Evaluation of Geometric-Uncertainty Alignment: We agree that standard OOD metrics including FPR95 and AUROC evaluate discriminative performance instead of direct metric correlation. For evidential deep learning, a practical criterion for geometry–uncertainty alignment is whether evidence magnitude reliably captures epistemic uncertainty across regions with different manifold densities.

  • Our controlled ablation between Ours (w/o Lali\mathcal{L}^{ali}) and Ours as shown in part of Table 9 demonstrates that Lali\mathcal{L}^{ali} systematically improves uncertainty separation, particularly rescuing sparse tail classes from severe miscalibration. This confirms that our threshold-based design effectively governs evidential confidence across diverse manifold regions.
  • We select the threshold-based design to preserve the hierarchical topology of the Lorentz manifold. That said, explicit functional regression under milder constraints is a promising complementary direction, which we reserve for future work (see the blue-highlighted text in Appendix E, Limitations and Future Work).
AuthorsOpenReview ↗

Responses to Reviewer KQ1M(Part 3)

Response to Question 2

We appreciate the reviewer raising this insightful point. To separate the OOD gains attributable to Lorentz geometry from those of the Lali\mathcal{L}^{ali}, we ran a controlled experiment. We reused the same evidence network, prototype design and training objective in standard Euclidean feature spaces, keeping all other experimental settings unchanged. Results on CIFAR100 and CIFAR100-LT (η=0.01\eta=0.01) are shown below:

Table 9: Ablation study for loss Lali\mathcal{L}^{ali} using CIFAR100 and CIFAR100‑LT (η=0.01\eta=0.01) as ID datasets. We additionally include a Euclidean counterpart equipped with its dedicated alignment loss Lali,e\mathcal{L}^{ali,e} for rigorous comparison. (Note: Due to web response length limits, C100 denotes CIFAR-100, and C100LT denotes CIFAR-100-LT (η=0.01\eta=0.01).)

MethodC100C100C100C100C100LTC100LTC100LTC100LT
OODOODOODIDOODOODOODID
FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$
EDL79.09 ± 0.8571.51 ± 0.4175.63 ± 0.5180.38 ± 0.1392.51 ± 0.6258.90 ± 0.6062.45 ± 0.7158.13 ± 0.24
EDL(w Lali,e\mathcal{L}^{ali,e})82.77 ± 0.0362.87 ± 0.0367.76 ± 0.0379.65 ± 0.0496.77 ± 0.0547.64 ± 0.1350.84 ± 0.2153.81 ± 0.28
Ours(w/o Lali\mathcal{L}^{ali})40.29 ± 1.1487.88 ± 0.5291.52 ± 0.1880.13 ± 0.1756.07 ± 1.7876.05 ± 1.0484.92 ± 0.7250.89 ± 0.46
Ours36.50 ± 2.0488.32 ± 0.6992.24 ± 0.4380.10 ± 0.1348.25 ± 3.3379.08 ± 1.0388.36 ± 0.9049.88 ± 0.62

Note on EDL (w/ Lali\mathcal{L}^{ali}): The extremely small standard deviations observed for EDL (w/ Lali\mathcal{L}^{ali}) stem from rounding values originally in the 10−410^{-4} magnitude (e.g., 0.0003463581241353642 for FPR95 and 0.00031084379783058355 for AUPR and 000343889819428845 for AUROC) to two decimal places, indicating a deterministic and highly stable convergence behavior under flat Euclidean geometry.

From these comparisons, we summarize the following observations:

  • Lorentz geometry delivers the main performance gain. Switching from Euclidean space to the Lorentz manifold without alignment loss (EDL vs. Ours (w/o Lali\mathcal{L}^{ali})) reduces OOD FPR95 by roughly half on CIFAR-100 (79.09%→40.29%79.09\% \rightarrow 40.29\%). This indicates the hyperbolic representation provides the core inductive bias for OOD detection.
  • The regularization term alone does not help in Euclidean space. Adding Lali,e\mathcal{L}^{ali,e} directly to Euclidean EDL degrades performance. Flat Euclidean space lacks the exponential representational capacity of hyperbolic manifolds; imposing evidence constraints within this linear space compresses features and distorts decision boundaries.
  • The two components work synergistically. Lali\mathcal{L}^{ali} brings further clear gains only when combined with Lorentz space. Negative curvature supplies enough representational capacity, so the alignment regularization can calibrate Dirichlet parameters without semantic collapse.

Taken together, the improvements arise from the synergy between Lorentz geometry and alignment regularization, where Lorentz representations form the basis. We have added these controlled results and the corresponding discussion to Table 9 in Appendix D.2 (see the blue-highlighted text).

Reviewer_w465OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LEDL, a Lorentz-space evidential learning framework for out-of-distribution detection. The method constructs Dirichlet parameters exclusively from the spatial components of Lorentz embeddings and introduces prototype-guided evidence alignment to regularize evidence magnitudes. Experiments demonstrate promising improvements, particularly on CIFAR100, CUB200, and long-tailed benchmarks.

Critical strengths

Overall, the method is a classic example of applying hyperbolic geometry to a downstream task or computational method. The topic is meaningful. Combining hyperbolic representation learning with EDL for OOD detection is a research direction worth exploring. The experimental results are reasonably competitive. For example, FPR95 on CIFAR100 drops from 47.79% for FEDL to 36.50%, and on CUB200 from 18.33% to 3.70%. There are also clear improvements on long-tailed CIFAR100. The experiments cover multiple classification datasets and report results over multiple random seeds. In addition, the method requires only a single forward pass and does not need to compute prototype distances at inference time, which gives it some practical value.

Critical weaknesses

Since z0=k+∥zs∥2z_0 = \sqrt{k + \lVert z_s \rVert^2} is fully determined by zsz_s, discarding z0z_0 loses no information, and the radial depth is still contained in ∥zs∥\lVert z_s \rVert. The "proof" in Appendix B.1 is invalid. Strategy-1 in Table 4 replaces the entire evidence head, so it cannot show that z0z_0 is harmful.

A Euclidean counterpart is missing. The EDL baselines differ from LEDL not only in geometry, as their evidence heads are different and they do not use Lali\mathcal{L}^{ali}.

Questions for authors

Could the authors move the preliminaries before the method section to give readers background on hyperbolic geometry and EDL?

What would the results be if the Euclidean features were fed directly into the subsequent computation in place of zsz_s, without the exponential map exp⁡0\exp_0? Since both lie in Rn\mathbb{R}^n, once the time component is dropped they could in theory be regarded as the same latent space.

Does "Ours" in Table 4 include Lali\mathcal{L}^{ali}? What about Strategy-1 and Strategy-2?

The baseline tables are taken from the original HEDL paper, but was HEDL itself reproduced by the authors?

6 comments
AuthorsOpenReview ↗

Responses to Reviewer w465 (Part 1)

To facilitate the reviewer in identifying which weakness or question is being addressed, we have organized the index numbers as follows:

Critical Weaknesses:

Weakness 1: Since z0=k+∥zs∥2z_0 = \sqrt{k + \lVert z_s \rVert^2} is fully ...

Weakness 2:A Euclidean counterpart ...

Questions For Authors:

Question 1: Could the authors ...

Question 2: What would the results ...

Question 3: Does "Ours" in Table 4 ...

Question 4: The baseline tables ...

To better answer weakness 1, I have split it into W1-1 and W1-2 for easier response.

W1-1: Since z0=k+∥zs∥2z_0 = \sqrt{k + \lVert z_s \rVert^2} is fully determined by , discarding loses no information, and the radial depth is still contained in ∥zs∥\lVert z_s \rVert . The "proof" in Appendix B.1 is invalid.

W1-2: Strategy-1 in Table 4 replaces the entire evidence head, so it cannot show that z0z_0 is harmful.

Response to Weakness 1-1

We sincerely thank the reviewer for their sharp and rigorous critique regarding Appendix B.1 and the theoretical treatment of the temporal coordinate z0z_0. We completely agree that from a strictly information-theoretic standpoint, because z0=k+∥zs∥2z_0 = \sqrt{k + \|z_s\|^2}, excluding z0z_0 from the evidence network does not mathematically delete the underlying radial magnitude, as it remains implicitly encoded within the norm ∥zs∥\|z_s\|.

We wish to clarify that our claim of preventing semantic contamination does not rely on an information-theoretic erasure or loss of the radial scale. Instead, it highlights a crucial operational distinction concerning gradient dynamics and evidence parameterization stability. When the raw time coordinate z0z_0—which inherently functions as a strictly positive, large common-mode scalar offset (k>0\sqrt{k} > 0)—is directly concatenated into the nonlinear evidence-generating network (as evaluated in Strategy-1), this large positive offset tends to dominate the feature activations and backpropagation gradients. Under skewed or heavy-tailed distribution shifts (such as CIFAR100-LT), this common-mode gradient dominance overwhelms fine-grained class-discriminative semantics.

By restricting the evidence parameterization strictly to the spatial components zsz_s, we successfully eliminate this gradient-dominating scale bias without discarding the underlying manifold geometry needed for distance measurement. We have thoroughly revised Appendix B.1(see the blue-highlighted text) in the updated manuscript to reframe the proof around optimization stability and gradient conditioning rather than information loss, thereby completely resolving any theoretical ambiguity.

AuthorsOpenReview ↗

Responses to Reviewer w465 (Part 2)

Response to Weakness 2

We sincerely thank the reviewer for this insightful comment and for identifying confounding factors for the Euclidean counterpart. To construct a rigorous ablation, we add a genuine Euclidean counterpart (EDL (w/ Lali,e\mathcal{L}^{ali,e})) to Table 9 in the revised appendix. In this configuration, class prototypes are updated with standard Euclidean exponential moving average (EMA), and distances are calculated by squared Euclidean metric. The backbone and optimization pipeline stay unchanged, without any hyperbolic-specific loss adaptation.

Table 9: Ablation study for loss Lali\mathcal{L}^{ali} using CIFAR100 and CIFAR100‑LT (η=0.01\eta=0.01) as ID datasets. We additionally include a Euclidean counterpart equipped with its dedicated alignment loss Lali,e\mathcal{L}^{ali,e} for rigorous comparison. (Note: Due to web response length limits, C100 denotes CIFAR-100, and C100LT denotes CIFAR-100-LT (η=0.01\eta=0.01).)

MethodC100C100C100C100C100LTC100LTC100LTC100LT
OODOODOODIDOODOODOODID
FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$
EDL79.09 ± 0.8571.51 ± 0.4175.63 ± 0.5180.38 ± 0.1392.51 ± 0.6258.90 ± 0.6062.45 ± 0.7158.13 ± 0.24
EDL (w Lali,e\mathcal{L}^{ali,e})82.77 ± 0.0362.87 ± 0.0367.76 ± 0.0379.65 ± 0.0496.77 ± 0.0547.64 ± 0.1350.84 ± 0.2153.81 ± 0.28
Ours(w/o Lali\mathcal{L}^{ali})40.29 ± 1.1487.88 ± 0.5291.52 ± 0.1880.13 ± 0.1756.07 ± 1.7876.05 ± 1.0484.92 ± 0.7250.89 ± 0.46
Ours36.50 ± 2.0488.32 ± 0.6992.24 ± 0.4380.10 ± 0.1348.25 ± 3.3379.08 ± 1.0388.36 ± 0.9049.88 ± 0.62

Note on EDL (w/ Lali,e\mathcal{L}^{ali,e}): The extremely small standard deviations observed for EDL (w/ Lali,e\mathcal{L}^{ali,e}) stem from rounding values originally in the 10−410^{-4} magnitude (e.g., 0.0003463581241353642 for FPR95 and 0.00031084379783058355 for AUPR and 000343889819428845 for AUROC) to two decimal places, indicating a deterministic and highly stable convergence behavior under flat Euclidean geometry.

As shown in Table 9, this Euclidean alignment variant still cannot overcome its intrinsic limitations. Restricted by the limited expressiveness and geometric property of flat space, this Euclidean baseline yields poor OOD performance. For instance, FPR95 stays at 82.77% on CIFAR100 and 96.77% on CIFAR100-LT. Flat Euclidean space lacks negative curvature and native manifold constraints. It cannot naturally mitigate feature overlap for complex distributions, and fails to suppress overconfident evidence extrapolation on sparse tail samples.

In comparison, our Lorentz framework integrates non-Euclidean hyperbolic geometry, spatial-only evidence parameterization and dynamic alignment, and obtains much stronger OOD results: FPR95 reaches 36.50% on CIFAR100 and 48.25% on CIFAR100-LT. This controlled comparison verifies that our performance improvement is not attributed merely to the alignment scheme or redesigned evidence head. Instead, it arises from the essential synergy between Lorentz manifold geometry and our alignment design, which alleviates feature crowding. We have extended the discussion of Table 9 in Appendix D.2 (see the blue-highlighted text) in the revised manuscript to clarify this distinction.

AuthorsOpenReview ↗

Responses to Reviewer w465 (Part 3)

Response to Question 1

We sincerely thank the reviewer for this constructive suggestion on manuscript organization. We fully agree that placing background content earlier improves readability greatly. In the revised manuscript, Section 3.1 is rearranged as an independent “Preliminaries” section immediately before the Methods chapter. We also expand the foundational background of hyperbolic geometry (Lorentz model) and Evidential Deep Learning (EDL) in this section, so that the introduction is more self-contained. Changes are marked in blue in the Section 3 Preliminaries.

Response to Question 2

We sincerely thank the reviewer for this insightful question. From a pure dimensional perspective, zsz_s and raw Euclidean features both reside in Rn\mathbb{R}^n after removing the time component. Despite sharing the same ambient dimension, features in our pipeline go through the exponential map exp⁡0k\exp_0^k and Lorentz transformation layer fLf^L. These operations fundamentally reshape their topological distribution and inner-product structure by introducing non-Euclidean geometric inductive bias.

Without exp⁡0k\exp_0^k, raw Euclidean features cannot be directly fed into subsequent Lorentz operations, as they lack the required hyperbolic geometry and Minkowski inner-product properties. Skipping this step would reduce the whole model to a standard Euclidean EDL pipeline.

To verify this difference empirically, we add a standard Euclidean EDL baseline under identical settings to Table 9 in the revised appendix. As shown in Table9, the Euclidean baseline(EDL) from Lorentz baseline (Ours(w/o Lali\mathcal{L}^{ali})) suffers substantial OOD degradation. For CIFAR100, FPR95 rises from our 40.29% to 79.09%; for CIFAR100-LT, it rises from 56.07% to 92.51%. Although vanilla EDL achieves higher ID accuracy under severe imbalance (CIFAR100-LT, η=0.01\eta=0.01: 58.13% versus our 49.88%), this gain comes from overconfident evidence extrapolation on sparse tail samples. Our Lorentz-based framework suppresses such overconfident extrapolation and prioritizes reliable uncertainty estimation for open-set rejection.

Reviewer_uxbjOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper extends Evidential Deep Learning to Lorentz hyperbolic space for open-set recognition and out-of-distribution detection. The experiments cover balanced, long-tailed, and fine-grained scenarios, and the appendix provides extensive theoretical discussion.

Critical strengths
  1. The paper points out that existing Evidential Deep Learning (EDL) methods and their variants are largely confined to Euclidean embedding spaces, making them ill-suited for handling visual features characterized by non-Euclidean geometry or hierarchical structures. To address this, the authors propose a Lorentz hyperbolic EDL framework.
  2. As shown in Table 3, the OOD detection metrics on the CIFAR100-LT dataset demonstrate significant improvement over methods like FEDL. This suggests that the combination of hyperbolic representations and evidential calibration may be particularly effective in scenarios involving long-tailed distributions or distribution shifts.
Critical weaknesses
  1. The prototypical deviation in Equation (8) employs the full Lorentz inner product, which includes the product of the temporal components. Although the evidence parameter α\alpha relies solely on spatial components, the selection of "top-s" large-deviation samples and the statistics of dynamic targets remain influenced by the temporal/radial geometry. This contradicts the core claim that "evidence is driven entirely by spatial semantics."
  2. In Table 2, the results for "Ours" on CIFAR100 are: FPR95 36.50, AUROC 92.24, and Acc 80.10. However, in Table 4, the same "Ours" method on CIFAR100 shows: FPR95 40.29, AUROC 91.52, and Acc 80.13. There is a clear inconsistency for the exact same full pipeline.
  3. In Table 3, for CIFAR100-LT (η=0.01\eta=0.01), the ID accuracy for "Ours" is 49.88, compared to 58.70 for FEDL—a drop of nearly 9 percentage points. While the OOD FPR95 shows significant improvement, open-set recognition should not come at the cost of a severe degradation in the classification of known classes.
  4. In Table 7, for the Flowers102 dataset, the FPR95 for "Ours" is 12.66±10.23, whereas it is 5.98±0.00* for HEDL and 8.84±0.19 for FEDL. "Ours" exhibits a worse mean and an extremely high variance. Although the performance on CUB200 is strong, this does not justify a claim of overall "solid robustness." An analysis of the source of the variance on the Flowers102 dataset is required.
Questions for authors
  1. Address the inconsistencies in the "Ours" results across Tables 2, 3, and 4, and ensure the values ​​are consistent throughout all tables.
  2. Discuss the trade-off between ID accuracy and OOD performance; specifically, clarify whether the significant drop in accuracy for CIFAR100-LT (η=0.01) is acceptable.
  3. Explain the source of the 10.23 standard deviation for the FPR95 metric on the Flowers102 dataset, and provide the results for each individual seed.
  4. Include a comparison with at least one hyperbolic OOD/OSR baseline, or provide a clear justification as to why such a comparison is not feasible.
4 comments
AuthorsOpenReview ↗

Responses to Reviewer uxbj (Part 1)

Response to Weakness 1

We sincerely thank the reviewer for this insightful comment. We agree that the full Lorentz inner product in Equation (8) includes the temporal component to maintain manifold geometry, and z0z_0 is mathematically coupled with the spatial norm under the hyperboloid constraint.

Table 4: Ablation on alternative Dirichlet parameters construction schemes from Lorentz representations, with CIFAR100 and CIFAR100‑LT (η=0.01\eta=0.01) as ID datasets. (Note: Due to web response length limits, C100 denotes CIFAR-100, and C100LT denotes CIFAR-100-LT (η=0.01\eta=0.01).)

MethodC100C100C100C100C100LTC100LTC100LTC100LT
OODOODOODIDOODOODOODID
FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$FPR95$\downarrow$AUPR$\uparrow$AUROC$\uparrow$Acc$\uparrow$
Ours (w/o Lali\mathcal{L}^{ali})40.29 ± 1.1487.88 ± 0.5291.52 ± 0.1880.13 ± 0.1756.07 ± 1.7876.05 ± 1.0484.92 ± 0.7250.89 ± 0.46
Strategy-151.72 ± 2.2086.74 ± 0.4989.45 ± 0.4580.13 ± 0.0688.60 ± 3.7262.59 ± 0.9067.36 ± 1.9536.19 ± 0.19
Strategy-238.50 ± 0.4487.86 ± 0.0891.70 ± 0.0980.22 ± 0.1264.53 ± 2.2769.92 ± 0.7381.54 ± 0.8937.79 ± 0.23

We would like to clarify that “evidence driven entirely by spatial semantics” describes the parameterization of Dirichlet evidence, instead of completely excluding radial geometry from the whole framework from an information-theoretic perspective. As shown in our ablation study (Strategy-1 in Table 4), directly feeding the full vector [z0,zs][z_0, z_s] into the nonlinear evidence network makes the strictly positive, lower-bounded time component (k>0\sqrt{k} > 0) introduce a large common-mode scalar offset. This offset dominates backpropagation gradients and masks fine-grained semantic differences under heavy-tailed distributions.

Our spatial-only parameterization is designed to remove numerical gradient dominance and common-mode bias in evidence mapping, while the full Lorentz inner product is still used to accurately measure geometric deviation. We have clarified this functional separation between evidence mapping and manifold metric in the revised Appendix B.1 (see the blue-highlighted text).

Response to Weakness 2 and Question 1

We sincerely thank the reviewer for this constructive suggestion on our naming convention, and we apologize for the confusion brought by the unclear description in the original manuscript.

We clarify that "Ours" in Tables 2, 3 and all other main tables consistently refers to our full pipeline. By contrast, Table 4 is an ablation table built to examine standalone hyperbolic evidence construction strategies (Strategy-1, Strategy-2, Spatial-only), excluding auxiliary objectives. For this reason, "Ours" in Table 4 corresponds to the baseline trained only with classification loss Lece\mathcal{L}^{ece} (w/o Lali\mathcal{L}^{ali}), so that we can fairly isolate the effect of geometric representations.

We have carefully inspected all tables in the paper to avoid further confusion. "Ours" in Tables 1, 2, 3, 6, 7 and 8 stands for the complete pipeline. Two revisions are made accordingly: (1) the label in Table 4 is updated to "Ours (w/o Lali\mathcal{L}^{ali})"; (2) the relevant text in Section 5.3 (Ablation Study) is revised to state this baseline setting clearly. All changes are marked in blue in the revised manuscript. We appreciate the reviewer’s careful checking.

AuthorsOpenReview ↗

Responses to Reviewer uxbj (Part 2)

Response to Weakness 3 and Question 2

We sincerely thank the reviewer for the constructive feedback on the trade-off between ID accuracy and OOD detection under severe long-tailed distributions (η=0.01\eta=0.01). We provide a detailed explanation based on our empirical results and design principle.

First, on balanced benchmarks including CIFAR-10 and CIFAR-100 in Table 1 and Table 2, our LEDL improves ID accuracy and OOD detection simultaneously. For example, on CIFAR-10, LEDL reaches 95.32% ID accuracy and reduces FPR95 to 17.54%. This joint improvement comes from a native property of hyperbolic space: its exponential volume expansion naturally widens inter-class margins, which strengthens closed-set boundaries and open-set separation without conflict.

Second, for the accuracy change under severe class imbalance (CIFAR100-LT with η=0.01\eta=0.01 in Table 3), the closed-set accuracy drop of our method (49.88% compared with FEDL’s 58.70%) is not accidental performance degradation. It is a necessary trade-off for reliable open-set recognition. Euclidean baselines such as FEDL tend to over-extrapolate and produce overconfident predictions for tail samples with limited features. Although this unrestrained extrapolation can boost closed-set accuracy by making risky guesses on rare classes, it breaks epistemic uncertainty calibration. The model will wrongly treat unseen OOD samples as known classes, as shown by FEDL’s poor FPR95 of 73.59% under η=0.01\eta=0.01.

By comparison, our LEDL framework uses spatial evidence parameterization and alignment loss Lali\mathcal{L}^{ali} to restrain overconfident evidence extrapolation on tail classes. It encourages the model to output higher, safer uncertainty for ambiguous or data-scarce inputs. As a result, our method sacrifices a small amount of closed-set accuracy for a large, meaningful improvement in OOD reliability, lowering FPR95 from FEDL’s 73.59% to 48.25%. In safety-critical real-world scenarios, preventing dangerous false classification of unknown anomalies matters more than squeezing minor overconfident accuracy gains from extremely rare tail samples. We hope this explanation addresses your concern about this practical trade-off.

Response Weakness 4 & Question 3

We sincerely thank the reviewer for carefully examining model robustness and the performance variance observed on the Flowers102 dataset. To guarantee full transparency and rule out concerns about result stability or selective reporting, we provide raw metrics from five independent random seeds (114, 654, 25, 281, 759) in the table below:

SeedFPR95 ↓\downarrowAUPR ↑\uparrowAUROC ↑\uparrowAcc ↑\uparrow
11410.8396.2398.0487.40
6545.8997.4298.8384.76
259.5496.5598.3087.62
2816.4497.1698.8397.36
75930.5992.6095.0586.60

We compute all aggregated metrics and standard deviations using sample variance with Bessel’s correction (N−1=4N-1=4).

The large variance mainly comes from seed 759, which yields an FPR95 of 30.59%. This behavior is tied to the nature of fine-grained geometric optimization. Flowers102 contains 102 visually similar classes with subtle inter-class distinctions and heavy feature overlap. For fine-grained tasks, some random initialization paths may fall into local solutions where decision boundaries of easily confused classes interfere with one another, raising the false-positive rate.

Baselines such as HEDL and FEDL show smaller variance on this small dataset because they adopt more conservative fitting in Euclidean space. However, their OOD performance drops substantially when transferred to larger, more complex fine-grained datasets like CUB200. On CUB200, our method maintains strong and stable results with FPR95 = 3.70 ±\pm 0.82, while HEDL and FEDL degrade noticeably. Excluding seed 759, our method delivers consistently good results on Flowers102, with FPR95 ranging from 5.89% to 10.83%. We appreciate the reviewer’s careful checking.

AuthorsOpenReview ↗

Responses to Reviewer uxbj (Part 3)

Response to Question 4

We sincerely thank the reviewer for this constructive suggestion and for recognizing the value of discussions on hyperbolic OOD detection.

Following this suggestion, we add comparisons with the recent state-of-the-art hyperbolic OOD work from Kasarla et al. [1], which is highly relevant to hyperbolic open-set recognition. Due to differences in original codebases and evaluation pipelines, we clearly state the experimental settings: Kasarla et al. adopt a heavier ResNet-34 backbone, while our LEDL obtains better results using the lighter ResNet-18. To match their far-OOD evaluation protocol covering MNIST, SVHN, Textures and Places365, we supplement MNIST in our test set to keep the benchmark suite fully consistent.

Table 10: Average OOD Performance Comparison with Hyperbolic OOD Detection Methods on CIFAR-100.} Performance metrics (FPR95$\downarrow$, AUPR$\uparrow$, AUROC$\uparrow$) are evaluated on far-OOD datasets (MNIST, SVHN, Textures, Places365). Baseline results are quoted from Kasarla et al. [1] using a ResNet-34 backbone, whereas our LEDL uses a lightweight ResNet-18.

FPR95 ↓\downarrowAUPR ↑\uparrowAUROC ↑\uparrow
MSP[2]49.46 ±\pm 0.2370.41 ±\pm 0.3982.43 ±\pm 0.26
TempScale[3]48.61 ±\pm 0.2471.13 ±\pm 0.4083.02 ±\pm 0.16
Odin[4]49.45 ±\pm 0.3470.26 ±\pm 0.9182.97 ±\pm 1.24
Gram[5]57.77 ±\pm 0.3164.64 ±\pm 5.1376.84 ±\pm 1.15
Energy[6]55.41 ±\pm 1.3061.83 ±\pm 5.0681.74 ±\pm 0.50
KNN[7]44.00 ±\pm 0.1073.71 ±\pm 1.4485.51 ±\pm 0.17
DICE[8]54.67 ±\pm 0.2466.35 ±\pm 3.7480.96 ±\pm 1.73
Rank Feat[9]49.91 ±\pm 0.3368.51 ±\pm 2.9881.25 ±\pm 0.35
ASH[10]55.29 ±\pm 0.0464.89 ±\pm 0.8676.83 ±\pm 0.51
SHE[11]53.78 ±\pm 0.2567.21 ±\pm 5.6682.02 ±\pm 0.60
GEN[12]48.70 ±\pm 0.3570.98 ±\pm 0.4382.95 ±\pm 0.40
NNGuide[13]57.94 ±\pm 1.0263.13 ±\pm 4.7781.23 ±\pm 0.99
SCALE[14]53.36 ±\pm 0.1566.71 ±\pm 1.1979.20 ±\pm 0.79
Ours44.18 ±\pm 1.7986.55 ±\pm 0.6189.06 ±\pm 0.56

Table10 summarizes the comparison. Kasarla et al. test multiple post-hoc distance scoring functions on hyperbolic prototypes, where their best KNN baseline achieves 44.00 FPR95, 73.71 AUPR and 85.51 AUROC. Our probabilistic evidential framework reaches competitive FPR95 at 44.18, and surpasses all listed baselines by a clear margin on AUPR (86.55) and AUROC (89.06). Notably, these improvements are obtained with the lighter ResNet-18 backbone, which demonstrates the higher efficiency and stronger representational capacity of our evidence-driven hyperbolic formulation. We have added this table and the corresponding discussion in Appendix D.3 Comparison with Hyperbolic OOD Methods in blue.

AuthorsOpenReview ↗

Responses to Reviewer uxbj (Part 4)

[1] Kasarla, T., van Spengler, M., & Mettes, P. (2026). Balanced hyperbolic embeddings are natural out-of-distribution detectors. International Journal of Computer Vision, 134(5).

[2] Hendrycks, D., & Gimpel, K. (2016). A baseline for detecting misclassified and out-of-distribution examples in neural networks [Preprint]. arXiv.

[3] Guo, C., Pleiss, G., Sun, Y., & Weinberger, K. Q. (2017). On calibration of modern neural networks [Preprint]. arXiv. [4] Liang, S., Li, Y., & Srikant, R. (2017). Enhancing the reliability of out-of-distribution image detection in neural networks [Preprint]. arXiv.

[5] Sastry, C. S., & Oore, S. (2020). Detecting out-of-distribution examples with Gram matrices. In International Conference on Machine Learning (pp. 8491–8501).

[6] Liu, W., Wang, X., Owens, J. D., & Li, Y. (2020). Energy-based out-of-distribution detection [Preprint]. arXiv.

[7] Sun, Y., Ming, Y., Zhu, X., & Li, Y. (2022). Out-of-distribution detection with deep nearest neighbors [Preprint]. arXiv.

[8] Sun, Y., & Li, Y. (2022). DICE: Leveraging sparsification for out-of-distribution detection. In Lecture notes in computer science (pp. 691–708).

[9] Song, Y., Sebe, N., & Wang, W. (2022). RankFeat: Rank-1 feature removal for out-of-distribution detection. In neural information processing systems (pp. 17885–17898).

[10] Djurisic, A., Bozanic, N., Ashok, A., & Liu, R. (2022). Extremely simple activation shaping for out-of-distribution detection [Preprint]. arXiv.

[11] Chen, Z., Huang, H.-Z., Deng, Z., & Wu, J. (2024). Shrinkage mamba relation network with out-of-distribution data augmentation for rotating machinery fault detection and localization under zero-faulty data. Mechanical Systems and Signal Processing, 224, 112145.

[12] Liu, X., Lochman, Y., & Zach, C. (2023). GEN: Pushing the limits of softmax-based out-of-distribution detection. In IEEE Computer Society Conference on Computer Vision and Pattern Recognition (pp. 23946–23955).

[13] Park, J., Jung, Y. G., & Teoh, A. B. J. (2023). Nearest neighbor guidance for out-of-distribution detection. In IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 1686–1695).

[14] Xu, K., Chen, R., Franchi, G., & Yao, A. (2023). Scaling for training time and post-hoc out-of-distribution detection enhancement [Preprint]. arXiv.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.