acceptodds
Under review as a conference paper at ICLR 2027

ZERO-AD: Learning Normality as a Zero-Level Set for Multivariate Time-Series Anomaly Detection

Abstract

Multivariate time-series anomaly detection (MTSAD) typically characterizes normal behavior indirectly, for example through reconstruction, forecasting, representation similarity, or normal prototypes. However, these objectives do not explicitly define the properties that normal temporal behavior should satisfy, and as a result, anomalous deviations can become entangled with ordinary modeling uncertainty. We introduce ZERO-AD, a geometric framework that models normal temporal behavior using a small set of learned implicit constraints, , such that normal windows satisfy . The common zero-level set of these constraints therefore provides a data-driven representation of normality, while anomalies are identified directly through violations of the learned constraints. To prevent degenerate solutions, ZERO-AD regularizes the local constraint geometry by encouraging complementary Jacobian directions, consistency across observed normal temporal variations, and finite-step sensitivity when moving away from the zero set. During inference, constraint violations are calibrated using held-out normal data and combined through reliability weighting, temporal persistence, and dense–sparse aggregation. ZERO-AD requires neither anomalous training samples nor synthetic anomaly generation, and it does not rely on a reconstruction decoder or forecasting head. On TSB-AD-M, ZERO-AD achieves the best aggregate performance on five of the six reported metrics, including a VUS-PR of 0.437 and a Point-F1 of 0.470, while using only approximately 0.07M parameters. These results demonstrate that directly learning the constraints satisfied by normal temporal behavior provides a compact and geometrically interpretable alternative for normal-only MTSAD.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time series, Anomaly Detection, Zero Level Set, Constraint Function

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

This work proposes a geometric framework that models normal temporal behavior using a compact set of learned implicit constraints. These constraints are learned exclusively from normal training samples, while anomalies are identified at inference time based on their violations of the learned constraints.

Critical strengths

The main strength of this work lies in its competitive experimental performance. Beyond this, however, I find it difficult to identify other particularly compelling strengths.

Critical weaknesses

I discuss the weaknesses one by one below, in descending order of importance.

  1. The novelty of this work is difficult to assess due to the unclear and, at times, unprofessional presentation of the methodology, which makes the authors’ core ideas hard to follow. First, the mathematical formulation is insufficiently rigorous and contains several apparent inconsistencies. For example, there seem to be potential errors or ambiguities in Eqs. (4)–(7), and I found the derivation particularly difficult to follow. The relationship between Δzl\Delta z_l and UU is not clearly explained. Similarly, the motivation for introducing the so-called temporal constraint network Fθ(Z)F_{\theta}(Z) is unclear, as is its specific role within the overall framework. More generally, similar issues appear throughout the methodology section. Many key concepts, mathematical relationships, and design motivations are introduced without sufficient explanation, making it difficult to understand the proposed method and to properly evaluate its technical novelty.

  2. The overall organization of the paper is rather unusual. In particular, the writing style of the Introduction is noticeably different from what I typically encounter in academic manuscripts. Based on my more than six years of reviewing experience, the highly formulaic and somewhat unnatural presentation raises concerns that the manuscript may have been heavily assisted by generative AI. This concern is further reinforced by Fig. 2, where several arrows exhibit visually implausible shapes and inconsistent structures.

  3. Although the experimental results appear competitive, the paper does not clearly explain how the proposed method exploits cross-variable dependencies in multivariate time series to improve anomaly detection.

  4. The visualization of temporal anomaly localization is rather limited, which makes it difficult to qualitatively assess and better understand the effectiveness of the proposed method. I suggest that the authors provide more comprehensive visualization results, following prior works [1,2], for example, by illustrating the localization performance on different types of anomalies, such as point anomalies and trend anomalies. Additionally, more most recently developed MTS analysis methods should also be included [3, 4].

[1] Anomaly transformer: Time series anomaly detection with association discrepancy. ICLR 2022.

[2] iTransformer: Inverted Transformers Are Effective for Time Series Forecasting. AAAI 2024.

[3] TimeFilter: Patch-Specific Spatial-Temporal Graph Filtration for Time Series Forecasting. ICML 2025.

[4] TimeMixer++: A General Time Series Pattern Machine for Universal Predictive Analysis. ICLR 2025.

Questions for authors

See weaknesses above.

7 comments
AuthorsOpenReview ↗

Response:1

We thank the reviewer for the constructive feedback. We clarify that Eqs. (4)–(7) are consistent with ZERO-AD's formulation. To keep the manuscript concise, some intermediate definitions and connections were presented compactly. We recognize that this made the relationship between input augmentation, the constraint map, and its Jacobian difficult to follow. We clarify these steps below without changing the objectives or results.

1. Eqs. (4)–(5): Input and constraint map. For a window Z=[z1,…,zL]⊤∈RL×DZ=[z_1,\ldots,z_L]^\top\in\mathbb{R}^{L\times D}, Eq. (4) defines Δz1=0\Delta z_1=0 and Δzℓ=zℓ−zℓ−1\Delta z_\ell=z_\ell-z_{\ell-1} for ℓ=2,…,L\ell=2,\ldots,L. Thus ΔZ∈RL×D\Delta Z\in\mathbb{R}^{L\times D} and U(Z)=[Z,ΔZ]∈RL×2DU(Z)=[Z,\Delta Z]\in\mathbb{R}^{L\times 2D}, with feature-wise concatenation. Writing the temporal network as gθg_\theta, the full mapping is

Fθ(Z)=gθ(U(Z))=[f1(Z),…,fK(Z)]⊤∈RK.F_\theta(Z)=g_\theta(U(Z))=[f_1(Z),\ldots,f_K(Z)]^\top\in\mathbb{R}^K.

Hence U(Z)U(Z) supplies observations and temporal changes; it is not another learned constraint map. Eq. (5),

Lzero=1K∥Fθ(Z)∥22,\mathcal{L}_{\mathrm{zero}}=\frac{1}{K}\lVert F_\theta(Z)\rVert_2^2,

encourages normal windows to satisfy the constraints, but alone permits the trivial solution Fθ≡0F_\theta\equiv0.

2. Eqs. (6)–(7): Jacobian and constraint geometry. The Jacobian is defined with respect to the original window:

Jθ(Z)=∂Fθ(Z)∂vec⁡(Z)∈RK×LD.J_\theta(Z)=\frac{\partial F_\theta(Z)}{\partial\operatorname{vec}(Z)}\in\mathbb{R}^{K\times LD}.

As U(Z)U(Z) is deterministic, the chain rule gives

Jθ(Z)=∂gθ(U)∂vec⁡(U)∣U=U(Z)∂vec⁡(U(Z))∂vec⁡(Z).J_\theta(Z)=\left.\frac{\partial g_\theta(U)}{\partial\operatorname{vec}(U)}\right|_{U=U(Z)}\frac{\partial\operatorname{vec}(U(Z))}{\partial\operatorname{vec}(Z)}.

Thus each Jacobian row is a constraint gradient over the full multivariate window, including dependence through ΔZ\Delta Z. Eq. (7),

Lorth=∥Jθ(Z)Jθ(Z)⊤−IK∥F2,\mathcal{L}_{\mathrm{orth}}=\lVert J_\theta(Z)J_\theta(Z)^\top-I_K\rVert_F^2,

encourages approximately unit-norm, mutually orthogonal local gradients, discouraging flat or redundant constraints. In particular, Fθ≡0F_\theta\equiv0 implies Jθ=0J_\theta=0 and a nonzero orthogonality penalty; no global orthogonality is claimed.

3. Temporal network and geometric objectives. The network in Section 3.2 is a lightweight differentiable parameterization of FθF_\theta, not a separate architectural novelty. Point-wise projection and channel mixing capture cross-variable interactions; dilated residual blocks capture temporal context; pooling and an MLP produce KK constraints. This compact design is useful because Jacobian regularization adds training cost. The other objectives are complementary: Eq. (8) penalizes ∥Jθ(Z)vec⁡(tZ)∥22\lVert J_\theta(Z)\operatorname{vec}(t_Z)\rVert_2^2 for a normalized transition between nearby normal windows; Eq. (9) preserves constraint satisfaction along a finite temporal step; and Eq. (10) encourages controlled finite-step growth in a sampled constraint-gradient direction. Together they shape sensitivity to departures while preserving observed normal temporal evolution.

4. Novelty, compactness, and evidence. We do not claim that zero-level sets or Jacobian regularization are individually new. ZERO-AD unifies (i) implicit normality constraints, (ii) complementary local departure geometry and temporal consistency, and (iii) detection through calibrated violations of those same constraints, without reconstruction or forecasting heads. This directly couples normality learning with anomaly evidence. Despite using only 0.07M parameters versus 0.3M for PaAno (approximately 4.3× fewer), ZERO-AD achieves the best aggregate results on five of six TSB-AD-M metrics (Table 1). Table 3 further confirms the contribution of geometric regularization through component ablations. These findings support competitive or improved detection performance with a substantially smaller model, although Jacobian-based training increases runtime.

We will revise Sections 3.2–3.3 to make the dimensions, composition, chain-rule derivation, and purposes of Eqs. (5)–(10) explicit, and clarify that the contribution is the geometry-aware constraint-learning framework rather than the backbone. We appreciate the reviewer's feedback on improving clarity.

AuthorsOpenReview ↗

Response: 2

We thank the reviewer for the constructive feedback regarding the manuscript's organization, writing style, and visual presentation. We acknowledge that certain aspects of the Introduction and Fig. 2 could be improved to ensure a more natural academic narrative and clearer technical presentation. These are primarily matters of exposition and figure design, which can be addressed directly in the revision.

Regarding generative AI assistance, we would respectfully clarify that its use was explicitly disclosed in the AI Use Statement of the original submission. As stated therein, generative AI tools were used for language-level assistance, including paraphrasing, rephrasing, and formatting, with the resulting content reviewed by the authors. The scientific contributions, mathematical formulation, methodological development, experimental design, implementation, reported results, and conclusions remain the responsibility of the authors. We recognize that some language-level revisions may have resulted in an overly structured writing style, and we will refine these passages to improve readability and academic presentation.

Regarding the introduction, we agree that its narrative can be made more direct and conventional. We will streamline the discussion by establishing a clearer progression from the limitations of existing MTSAD approaches to the motivation for learning implicit normality constraints, followed by the geometric formulation and principal contributions of ZERO-AD. We will also remove unnecessary rhetorical transitions and repetitive phrasing. These changes will improve the clarity of the scientific motivation without altering the proposed method.

Regarding Fig. 2, we appreciate the reviewer's observation about the arrow alignment and visual consistency. The figure was designed to summarize the complete ZERO-AD framework, including constraint learning, geometric regularization, calibration, and anomaly scoring, within a compact layout to accommodate the manuscript's space constraints. Consequently, some connectors and arrows were not optimally aligned. We agree that these minor presentation details can be improved and will provide a revised figure with consistent arrowheads, properly aligned connectors, and a clearer visual arrangement of the existing components. This refinement will improve readability while preserving the same computational flow and technical content.

We appreciate the reviewer bringing these presentation issues to our attention. We will address them through careful editing and figure refinement. Importantly, these revisions do not require changes to the proposed mathematical framework, experimental methodology, or reported results. We believe the improved presentation will allow the technical contributions and empirical findings to be assessed more clearly on their scientific merits.

AuthorsOpenReview ↗

Response: 3

We thank the reviewer for this observation and the opportunity to clarify the role of cross-variable dependencies in ZERO-AD. A key aspect of our framework is that normality is characterized through joint constraints over the entire multivariate temporal window, rather than independent per-variable modeling. This allows ZERO-AD to represent relationships among variables and identify anomalies arising from violations of learned joint behavior. Below, we explain the underlying mechanism and present additional controlled ablation experiments demonstrating its contribution to detection performance.

  1. Joint multivariate constraint learning. ZERO-AD does not learn separate constraints for individual variables. Instead, each constraint is defined over the complete multivariate window Z∈RL×DZ\in\mathbb{R}^{L\times D}. The augmented input is

U(Z)=[Z,ΔZ]∈RL×2D,U(Z)=[Z,\Delta Z]\in\mathbb{R}^{L\times 2D},

combining all variables and their temporal differences. The initial point-wise projection jointly maps these features into a shared hidden representation. Dilated depth-wise convolutions capture temporal patterns, while gated point-wise channel mixing enables interactions across learned features. The resulting representation is mapped to

Fθ(Z)=[f1(Z),…,fK(Z)]⊤∈RK.F_\theta(Z)=[f_1(Z),\ldots,f_K(Z)]^\top\in\mathbb{R}^K.

Thus, each fk(Z)f_k(Z) can depend jointly on multiple variables and their temporal evolution, rather than being restricted to an individual channel. This enables ZERO-AD to represent cross-variable relationships without requiring a predefined dependency graph.

  1. Cross-variable dependencies in the learned geometry. The geometric formulation also operates on the complete multivariate window. Equation (6) defines

Jθ(Z)=∂Fθ(Z)∂vec⁡(Z)∈RK×LD.J_\theta(Z)=\frac{\partial F_\theta(Z)}{\partial\operatorname{vec}(Z)}\in\mathbb{R}^{K\times LD}.

Each Jacobian row characterizes the sensitivity of a constraint across the temporal-variable coordinates. Jacobian orthogonality encourages complementary local departure directions, while temporal consistency encourages invariance along observed normal evolution.

For example, two variables may normally satisfy a lagged relationship, xt(2)≈axt−τ(1)+bx_t^{(2)}\approx a x_{t-\tau}^{(1)}+b. Both may remain individually within normal ranges while their relationship becomes anomalous. A learned joint constraint can respond to such a relational violation, which independent marginal modeling may overlook. This illustrates the type of dependency ZERO-AD can learn implicitly.

  1. Additional cross-variable ablation experiments. To evaluate this contribution directly, we conducted three controlled ablations following the TSB-AD-M evaluation protocol: (i) Dependency-destroyed, which disrupts cross-variable alignment while preserving channel-wise marginal and temporal behavior; (ii) Channel-independent ZERO-AD, which learns constraints separately for each variable and aggregates their scores; and (iii) Without cross-channel mixing, which removes gated point-wise mixing within the temporal blocks while retaining the initial multivariate projection.

The experimental results are summarized below:

VariantVUS-PRVUS-ROCRange-F1AUC-PRAUC-ROCPoint-F1
Dependency-destroyed0.3790.7310.3720.3350.7240.418
Channel-independent ZERO-AD0.4060.7650.3910.3520.7550.421
w/o cross-channel mixing0.4180.7710.4110.3640.7630.441
Full ZERO-AD0.4370.7800.4350.3960.7820.470

Full ZERO-AD outperforms all three variants across all six metrics. Compared with channel-independent learning, VUS-PR improves from 0.406 to 0.437 and Point-F1 from 0.421 to 0.470, supporting the benefit of learning joint multivariate constraints. Removing gated cross-channel mixing reduces VUS-PR to 0.418, indicating the contribution of feature interactions within the temporal blocks. Disrupting cross-variable alignment produces the largest VUS-PR degradation (0.437 to 0.379), further highlighting the importance of preserving multivariate relationships.

These results provide direct empirical support that ZERO-AD's detection performance benefits from joint multivariate modeling rather than solely independent per-variable temporal patterns.

We will incorporate these additional cross-variable ablation results into the revised manuscript, either in the main ablation study (Section 5.2) or in the Appendix, depending on space availability.

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

This paper proposes ZERO-AD, a normal-only multivariate time-series anomaly detection method that learns a set of implicit constraints to characterize normal behavior. Anomalies are detected through violations of these constraints, rather than reconstruction or forecasting errors. The method introduces geometric regularization to prevent trivial solutions and preserve normal temporal variations. Experiments on TSB-AD-M show competitive performance, supported by ablation studies, parameter sensitivity analyses, and failure-case discussions.

Critical strengths
  1. The motivation is reasonable and intuitive. The method learns stable patterns in multivariate time series and identifies anomalies as deviations from these patterns. Compared with reconstruction- or forecasting-based methods, which introduce specific inductive biases, I find this formulation cleaner and more direct.

  2. The model is relatively simple and easy to understand. The ablation studies and sensitivity analyses, particularly those on the number of constraints K and window size, are helpful in understanding the main contributions.

  3. The paper also includes failure-case analyses, which I believe are valuable to the TSAD community.

Critical weaknesses
  1. The experiments focus exclusively on TSB-AD-M, without evaluating univariate time series. I understand that the method may benefit from modeling relationships between variables, while potentially facing challenges on univariate data, especially strongly periodic series. Still, I would like to see some discussion of its applicability to univariate TSAD. In particular, can the learned constraints capture interpretable patterns such as periodicity, trends, and noise?
Questions for authors

See weakness.

4 comments
AuthorsOpenReview ↗

Response: 1

We thank the reviewer for this insightful question and the opportunity to further demonstrate ZERO-AD's applicability to univariate time series. We emphasize that ZERO-AD's fundamental learning principle does not depend on cross-variable relationships. Rather, it learns implicit constraints characterizing normal temporal behavior, which remain meaningful even for a single variable. To substantiate this, we conducted additional experiments on the univariate TSB-AD benchmark, complementary cross-variable ablations, and synthetic analyses of periodicity, trends, and noise.

1. Applicability to univariate time series. ZERO-AD naturally extends to univariate TSAD by setting D=1D=1. A temporal window becomes Z∈RL×1Z\in\mathbb{R}^{L\times1}, with augmented input

U(Z)=[Z,ΔZ]∈RL×2.U(Z)=[Z,\Delta Z]\in\mathbb{R}^{L\times2}.

The network learns KK implicit temporal constraints,

Fθ(Z)=[f1(Z),…,fK(Z)]⊤∈RK,F_\theta(Z)=[f_1(Z),\ldots,f_K(Z)]^\top\in\mathbb{R}^K,

with Jacobian

Jθ(Z)=∂Fθ(Z)∂vec⁡(Z)∈RK×L.J_\theta(Z)=\frac{\partial F_\theta(Z)}{\partial\operatorname{vec}(Z)}\in\mathbb{R}^{K\times L}.

The geometric objectives remain applicable: Jacobian orthogonality encourages complementary temporal departure directions, while temporal tangent and finite-step consistency preserve observed normal evolution. Thus, in the univariate setting, ZERO-AD learns relationships among observations across time rather than relationships among variables, without requiring changes to its fundamental formulation.

2. Additional univariate benchmark results. To directly address the reviewer's concern, we evaluated ZERO-AD on TSB-AD-U, following the univariate benchmark setting used by PaAno (ICLR 2026), with the same six evaluation metrics.

MethodVUS-PRVUS-ROCRange-F1AUC-PRAUC-ROCPoint-F1
PaAno0.5200.8900.4800.4600.8600.510
ZERO-AD0.5510.8740.5380.4850.8570.542

ZERO-AD outperforms PaAno on four of six metrics, including the primary VUS-PR metric (0.551 vs. 0.520), Range-F1 (0.538 vs. 0.480), AUC-PR (0.485 vs. 0.460), and Point-F1 (0.542 vs. 0.510). VUS-ROC and AUC-ROC are slightly lower. These results demonstrate that the learned temporal constraints remain effective without cross-variable information, providing direct empirical support for ZERO-AD's applicability to univariate anomaly detection.

3. Complementary evidence from cross-variable ablations. To further distinguish temporal constraint learning from cross-variable dependency modeling, we conducted controlled ablations on TSB-AD-M.

VariantVUS-PRVUS-ROCRange-F1AUC-PRAUC-ROCPoint-F1
Dependency-destroyed0.3790.7310.3720.3350.7240.418
Channel-independent0.4060.7650.3910.3520.7550.421
Full ZERO-AD0.4370.7800.4350.3960.7820.470

The full model outperforms both variants across all six metrics. These findings establish an important distinction: cross-variable dependencies provide additional benefits in multivariate settings, but are not a prerequisite for learning informative temporal normality constraints. The univariate benchmark results independently demonstrate the effectiveness of the underlying constraint-learning principle.

4. Understanding periodicity, trends, and noise. The synthetic visualizations examine representative periodic, trending, and noisy signals, together with their learned constraint responses, constraint-violation patterns, and final anomaly scores. These analyses provide a qualitative understanding of how ZERO-AD captures normal temporal characteristics and responds to deviations, including changes in strongly periodic behavior.

Importantly, ZERO-AD learns these temporal characteristics implicitly, without requiring predefined periodic, trend, or noise decomposition modules. Although individual constraints are not necessarily associated with a unique temporal property, their responses provide interpretable evidence of departures from learned normal behavior.

Together with the improved TSB-AD-U benchmark results, these synthetic analyses provide complementary quantitative and qualitative evidence supporting ZERO-AD's applicability to diverse univariate temporal patterns.

We will incorporate the additional univariate benchmark results, complementary cross-variable ablations, and synthetic visualizations into the revised manuscript, either in the main text or Appendix, depending on space availability.

AuthorsOpenReview ↗

Visualization plots

We have included additional visualization plots in Appendix G of the revised manuscript, using controlled synthetic datasets to demonstrate how the learned constraints capture periodicity, trend, and noise characteristics of normal temporal behavior and how deviations from these patterns produce informative constraint violations for anomaly detection.

Thank you for the detailed response. The authors have addressed most of my concerns, and I will therefore maintain my positive scores. However, I agree with Reviewer neDe that the presentation of the paper still needs improvement, particularly in terms of clearly motivating and explaining the proposed method. I encourage the authors to address these issues in the final version.

We sincerely thank the reviewer for the positive assessment and constructive feedback. In response to the reviewers' suggestions, we have revised the manuscript, particularly the Introduction and Methodology sections, to more clearly highlight the motivation, core intuition, and technical details of ZERO-AD. We will continue refining the presentation in the final version to make the motivation and methodology even clearer and more accessible. We greatly appreciate the reviewer's valuable feedback and continued support.

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

This paper proposes ZERO-AD, which uses multiple relation functions whose values are expected to be close to zero for normal windows. To prevent degenerate solutions, the method combines five loss terms (Lzero\mathcal{L}{zero}, Lorth\mathcal{L}{orth}, Ltan\mathcal{L}{tan}, Ldist\mathcal{L}{dist}, and Ltz\mathcal{L}_{tz}). The reported experiments compare ZERO-AD with PaAno and show improvements in aggregate performance.

Critical strengths
  1. Using multiple aspects of relation functions: The idea of using relation functions to characterize normal windows is interesting. The authors construct a multi-objective loss function to capture different properties of these relations.
  2. Failure-case analysis: The proposed approach performs poorly against PaAno in several cases. Analyzing these failures is useful for understanding the limitations of the method.
Critical weaknesses
  1. Method complexity: The method uses five loss terms to learn the geometry, and calculating anomaly scores requires additional calibration. The referee feels that the authors have added several components one by one to improve performance, without sufficiently explaining why they are all needed as part of a unified approach. Although the paper reports the final hyperparameter settings, it does not clearly explain how they were selected. This makes it difficult to assess the tuning burden of the complete method.
  2. Performance relative to complexity: The failure-case analysis is useful, and ZERO-AD achieves the best aggregate results on five of the six metrics. However, the referee is not convinced that the performance improvements justify the added complexity of the training and scoring procedures.
Questions for authors
  1. How were the loss weights and scoring hyperparameters selected using the 20-series tuning set? Please clarify the tuning procedure and the effort required.
  2. Given the modest improvement over PaAno and the mixed results across datasets, how do the authors justify the added complexity of training and scoring? Can comparisons with a simpler, appropriately tuned alternative demonstrate the need for the full method?
6 comments
AuthorsOpenReview ↗

Response: 1

We thank the reviewer for the constructive feedback and the opportunity to clarify the methodological design and hyperparameter-selection procedure. We would like to emphasize that the five loss terms in ZERO-AD are designed to satisfy three complementary geometric requirements, rather than functioning as independently introduced mechanisms. Their mathematical roles, empirical contributions, and the tuning protocol are clarified below.

1. Unified motivation for the five terms. The complete training objective is defined in Eq. (12). Its components serve three related geometric requirements:

(i) Normality constraint satisfaction: L_zero encourages normal windows to satisfy the learned zero-level constraints. However, this objective alone permits the trivial constant-zero solution, making geometric regularization necessary.

(ii) Non-degenerate local geometry and temporal consistency: L_orth discourages flat or redundant constraints by encouraging approximately orthonormal Jacobian gradients, while L_tan encourages observed normal temporal transitions to remain tangent to the learned zero set.

(iii) Finite-step geometric consistency: L_tz extends temporal tangent preservation to finite displacements, while L_dist encourages controlled constraint growth along local departure directions. Thus, L_tan and L_tz enforce temporal consistency at differential and finite scales, whereas L_orth and L_dist provide complementary local and finite-step control of departure geometry.

These terms support a single geometric principle: normal temporal evolution should remain close to the learned zero set, while departures from normality should produce informative constraint violations. This interpretation is consistent with the local geometric analysis in Appendix B.1.

2. Empirical justification through ablations. The submitted Table 3 demonstrates the contribution of the geometric components:

VariantVUS-PRAUC-PR
Full ZERO-AD0.4370.396
w/o orthogonality0.3930.330
w/o temporal consistency0.3900.330
w/o distance consistency0.3950.333
w/o zero-set geometry0.3940.328

Removing each component degrades performance. In particular, removing the geometric regularizers while retaining normality fitting substantially reduces performance, demonstrating that simply minimizing constraint values is insufficient. These ablations support the complementary contributions of the geometric objectives rather than their inclusion as arbitrary performance-enhancing terms.

3. Role of calibration. Calibration serves a distinct purpose from geometric learning. Since different learned constraints can exhibit different normal response scales and variability, robust normalization is necessary before aggregating their violations. ZERO-AD estimates median/MAD statistics from held-out normal data, followed by reliability weighting and temporal dense–sparse aggregation. This procedure is deterministic and requires no additional neural network, anomaly-label supervision, gradient optimization, or test-time adaptation. Thus, calibration is not another independently trained component but a normal-only normalization and scoring procedure.

4. Hyperparameter selection and tuning burden. Following the TSB-AD-M evaluation protocol adopted by PaAno (ICLR 2026), 20 designated series are used for hyperparameter tuning and the remaining 180 for final evaluation. ZERO-AD uses a single global configuration across all evaluation series, without dataset-specific hyperparameter tuning. The submitted configuration uses L=64, H=64, K=8, with loss weights (1.0, 0.10, 0.10, 0.05, 0.10), corresponding to the five terms in Eq. (12).

Importantly, the parameter-sensitivity experiments in Fig. 4 are post-hoc robustness analyses, not additional tuning on the evaluation series. They show a favorable performance region around the selected configuration rather than dependence on an isolated parameter setting. All configurations were fixed before final evaluation, without using anomaly labels from the evaluation series for hyperparameter selection.

5. Model complexity and efficiency. Despite the geometry-aware objectives, ZERO-AD uses only approximately 0.07M parameters, compared with 0.3M for PaAno, while achieving the best aggregate performance on five of six TSB-AD-M metrics (Table 1). Although Jacobian regularization increases training computation, this overhead does not apply during inference, which requires only a forward pass and deterministic score aggregation.

Overall, ZERO-AD combines a compact constraint model, mathematically connected geometric regularizers, and deterministic normal-only calibration within a unified framework. The submitted ablation and sensitivity studies support the contribution and robustness of these components.

AuthorsOpenReview ↗

Response: 2

We thank the reviewer for the constructive feedback and the opportunity to clarify ZERO-AD's performance–complexity trade-off. We emphasize that ZERO-AD achieves improved anomaly-detection performance with a substantially smaller model, while its additional geometric computation is concentrated in training rather than inference. The following quantitative results further substantiate the effectiveness and practical efficiency of this design.

1. Model compactness and computational efficiency. ZERO-AD contains only 0.07M parameters, compared with 0.3M for PaAno (approximately 4.3× fewer). It requires no reconstruction decoder, forecasting head, auxiliary prediction network, or test-time optimization. Jacobian-based geometric regularization is performed during training, with its terms evaluated once every four minibatches to control computational overhead.

To further clarify the computational requirements, we decomposed the reported 19.5 s runtime into its principal components:

ComponentRuntimeShare
Standard training7.45 s38.2%
Jacobian-based geometry optimization11.96 s61.3%
Calibration0.03 s0.15%
Inference + scoring0.07 s0.36%
Total (rounded)19.50 s100%

Approximately 99.5% of the runtime is associated with training, whereas calibration and inference together require only about 0.5%. Importantly, no Jacobian computation is required during inference. Once trained, ZERO-AD performs a single forward pass followed by deterministic aggregation of only K=8 constraint responses. Thus, the geometry-aware optimization does not introduce additional Jacobian computation or optimization into deployment.

2. Performance relative to model complexity. As reported in Table 1, ZERO-AD achieves the best aggregate performance on five of six TSB-AD-M metrics, despite using substantially fewer parameters than PaAno. Improvements include VUS-PR (0.437 vs. 0.430), Range-F1 (0.435 vs. 0.410), AUC-PR (0.396 vs. 0.380), AUC-ROC (0.782 vs. 0.760), and Point-F1 (0.470 vs. 0.430).

Importantly, aggregate results also conceal substantial dataset-level gains (Table 2). On Exathlon, VUS-PR improves from 0.777 to 0.932 and Range-F1 from 0.605 to 0.811. On Genesis, VUS-PR increases from 0.296 to 0.678, while SWaT shows improvements in AUC-PR from 0.217 to 0.536 and Point-F1 from 0.321 to 0.663. These results demonstrate meaningful improvements across diverse temporal systems beyond the aggregate comparison.

3. Experimental justification for geometric learning and calibrated scoring. Appendix D.3 (Table 7) directly separates the contributions of geometry-aware learning and score aggregation:

VariantVUS-PRRange-F1
Zero-only + Simple Score0.3550.330
Zero-only + Full Score0.3940.385
Full Geometry + Simple Score0.3970.378
Full ZERO-AD0.4370.435

These controlled results demonstrate that the improvement is not merely attributable to a more elaborate scoring procedure. Introducing geometric learning alone increases VUS-PR from 0.355 to 0.397, while calibrated scoring alone improves it to 0.394. Combining both achieves 0.437, with a corresponding Range-F1 improvement from 0.330 to 0.435.

This establishes the complementary contributions of both stages: geometry-aware training learns informative normality constraints, while normal-only calibration and aggregation improve the reliability of their anomaly evidence. The improvement is therefore supported by explicit ablations rather than increased architectural capacity or scoring complexity alone.

4. Overall performance–complexity trade-off. ZERO-AD deliberately invests additional computation during training to learn structured normality constraints, while maintaining a compact architecture and simple inference procedure. Its approximately 4.3× parameter reduction, improvements across five aggregate metrics, substantial dataset-level gains, and controlled geometry-versus-scoring ablations collectively support the effectiveness of this design.

We will include the detailed runtime decomposition and corresponding performance–complexity analysis in the revised Appendix.

AuthorsOpenReview ↗

Response: 3

We thank the reviewer for the constructive feedback. Our approach follows the TSB-AD-M benchmark protocol adopted by PaAno (ICLR 2026), using 20 designated series for hyperparameter tuning and 180 held-out series for final evaluation. We employ a lightweight sequential search using VUS-PR as the selection metric, avoiding an exhaustive grid search over all parameter combinations.

1. Sequential hyperparameter selection. We evaluate a small set of candidate values for one parameter at a time while keeping the remaining parameters fixed. The candidate achieving the highest aggregate VUS-PR across the 20 tuning series is retained before proceeding to the next parameter. The principal candidate values are summarized below:

ParameterCandidate valuesSelected
Number of constraints K2, 4, 8, 12, 168
Window length L32, 48, 64, 96, 12864
Orthogonality weight0.05, 0.10, 0.15, 0.200.10
Temporal consistency weight0.05, 0.10, 0.15, 0.200.10
Distance consistency weight0.05, 0.10, 0.15, 0.200.05
Temporal pooling weight0, 0.10, 0.25, 0.50, 0.750.25
Sparse-score weight0, 0.05, 0.15, 0.30, 0.500.15

The selected geometric regularization weights are 0.10 for orthogonality, 0.10 for temporal tangent consistency, 0.05 for distance consistency, and 0.10 for finite-step temporal consistency, as reported in Appendix C (Table 4). The remaining calibration settings are fixed consistently across datasets.

2. Separation of tuning and final evaluation. Following the same benchmark-level tuning and per-series normal-training setting as PaAno, each series uses its predefined normal training prefix for model learning. ZERO-AD further partitions this prefix chronologically into 70% fitting, 15% validation, and 15% calibration.

During hyperparameter selection, candidate configurations are compared using VUS-PR exclusively on the 20 designated tuning series. Once selected, the global configuration is frozen and applied unchanged to all 180 held-out evaluation series.

For each evaluation series, model fitting, checkpoint selection, and calibration are performed using only its normal training prefix. The 180 evaluation series are never used to select or modify global hyperparameters, and their anomaly labels and evaluation metrics are used exclusively for final benchmark assessment. This clearly separates hyperparameter selection from normal-only model learning and final evaluation.

3. Tuning effort and robustness. The sequential strategy avoids the combinatorial cost of exhaustive grid search, and global tuning is performed only once rather than repeated for each evaluation series. Median/MAD calibration statistics and constraint reliability estimates are computed directly from held-out normal calibration data, without additional optimization.

The parameter-sensitivity analyses in Fig. 4 are post-hoc robustness experiments, separate from the original tuning procedure. They demonstrate a favorable performance region around the selected configuration, indicating that ZERO-AD does not depend on narrowly isolated hyperparameter values.

Overall, ZERO-AD follows the established TSB-AD-M protocol used by PaAno, with a computationally manageable sequential search and a single global configuration across all 180 evaluation series.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers and AC: Summary of paper revisions and Rebuttal

We sincerely thank all reviewers for their constructive feedback.

Reviewers TxVw and yDQf indicated that the rebuttal addressed their concerns, and we have incorporated the requested additional experiments, ablations, visualizations, and analyses into the current revised manuscript, primarily in Appendix G. These additions are intended to provide further evidence and clarification for the originally proposed ZERO-AD framework.

Reviewer neDe noted in the initial follow-up that our response had partially addressed the concerns, with the remaining issue focused mainly on the motivation in the Introduction and the clarity of the Methodology. At that stage, the substantially revised presentation had not yet been submitted. We have since refined the Introduction to make the motivation and central research question more explicit, and reorganized the Methodology with a clearer flow, explicit notation definitions, and concise explanations of the role of each component. Key motivational and methodological statements are also highlighted to make the paper easier to follow.

Importantly, these revisions do not change the proposed method, mathematical formulation, training objective, evaluation protocol, or conclusions of ZERO-AD. They clarify the original formulation and add supporting experiments requested during review.

For the current revised submission, all additional results and analyses are included in Appendix G so that they are directly available to the reviewers and Area Chair. In the final version, we will appropriately organize the most important supporting results between the main paper and appendix for clarity and space, and incorporate any remaining presentation clarification suggested during the discussion.

We hope the revised manuscript now addresses the remaining presentation concern raised by Reviewer neDe, and we respectfully request the reviewer and Area Chair to consider the revised manuscript in the final assessment.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.