acceptodds
Under review as a conference paper at ICLR 2027

CARE: A Cascaded Framework for Efficient and Reliable Time Series Anomaly Detection

Abstract

Deep learning models have achieved state-of-the-art performance in time series anomaly detection, yet their complex architectures incur substantial inference overhead. Existing methods uniformly apply costly inference to all inputs, ignoring that anomalies are inherently scarce and most temporal data exhibit predictable patterns. To address this inefficiency, we propose CARE, a model-agnostic cascaded inference framework that pairs a Lightweight Pre-filter Model (LPM) with a high-capacity Complex Detection Model (CDM). The LPM rapidly filters high-confidence normal samples using a Residual MLP-based AutoEncoder and a Normality-Conditioned Gating mechanism. The gating mechanism incorporates Structure Attention and Risk-State Memory to capture channel-wise anomaly contributions and temporal risk evolution. A budget-aware gating objective learns normality confidence rankings while aligning CDM allocation with the available computational budget. Extensive experiments across eight real-world benchmarks demonstrate that CARE effectively isolates high-confidence normal samples. By prioritizing low-confidence samples for CDM processing, our framework achieves inference speedups of up to compared to the most accurate SOTA approaches, while maintaining competitive detection quality.

Primary area:
learning on time series and dynamical systems
Keywords:
Anomaly detection, Times series, Cascade Inference

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Fz4MOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CARE, a model-agnostic cascaded framework for improving the inference efficiency of time series anomaly detection while maintaining detection performance. CARE combines a Lightweight Pre-filter Model (LPM) with a pretrained Complex Detection Model (CDM). The LPM employs a Normality-Conditioned Gating mechanism that integrates channel-wise Structure Attention and temporal Risk-State Memory to estimate normality confidence and selectively route uncertain samples to the CDM. The gating network is trained using relative rankings of CDM anomaly scores through a budget-aware objective, without requiring ground-truth anomaly labels. Additionally, a distribution-based score alignment strategy is introduced to reconcile the outputs of the two models.

Critical strengths
  1. The motivation is convincing, as many time series do not necessarily require complex models for effective anomaly detection.

  2. Although the proposed framework is relatively straightforward, I believe its design is valuable for real-world applications and efficient deployment.

Critical weaknesses
  1. The framework combines several design choices, including window/point-level errors, Structure Attention, and multiple weighted loss terms. However, their individual contributions are not sufficiently investigated.

    • (a) The window-level and point-level errors are not fully ablated, although this is a relatively minor concern.
    • (b) Several ablation results are very close to those of the full model, raising concerns about their statistical significance. Reporting results across multiple seeds would strengthen the conclusions. Additionally, the anonymous repository does not appear to provide complete code for reproducing these ablation studies.
  2. The experiments lack comparisons with some stronger recent TSAD methods. I suggest considering representative baselines from the TSB-AD benchmark to better establish the competitiveness of the proposed framework.

  3. Related to the previous point, the choice of complex models may influence the conclusions. Evaluating stronger and more diverse models could provide further insights into the effectiveness and generalizability of the cascading strategy.

Questions for authors

My main concerns are the selection of baselines and the marginal performance differences observed in several ablation studies. Including 1–2 stronger baselines and providing the complete code for the ablation studies would substantially increase my confidence in the results. If I can successfully reproduce and verify the reported findings, I would be willing to raise my rating.

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

The paper proposes CARE, a cascaded inference framework for efficient time series anomaly detection (TSAD). CARE combines a lightweight pre-filter model (LPM) with a high-capacity Complex Detection Model (CDM). A budget-aware gating objective learns normality confidence rankings while aligning CDM allocation with the available computational budget.

Critical strengths
  1. The performance–efficiency trade-off in deep TSAD is real and underexplored. Most TSAD papers optimize accuracy with uniform complex inference.
  2. CARE wraps any pretrained CDM without modifying it, and the paper demonstrates this across four CDM backbones.
Critical weaknesses
  1. The compared baselines are limited, more related works in 2026 should be included.
  2. Although the inference time has been shortened, a more complex two-stage training process has been introduced, which leads to an increase in training time.
  3. The NCG is trained using CDM scores on the training set as pseudo-labels. If the CDM systematically over- or under-scores certain patterns, the gate inherits those biases.
  4. Some of the content is overly repetitive. For instance, the module introduction in Section 3.1 and most of the subsequent chapters overlap with each other.
Questions for authors

Please see Weaknesses.

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

The paper argues that existing time series anomaly detection methods face a large trade-off between computational cost and detection accuracy. To address this, it proposes CARE, which uses an accurate but expensive model only when needed and otherwise relies on a method with short inference time and reasonable performance. CARE first processes all data with a cheap model, an autoencoder-based anomaly detector. Three types of reconstruction errors from the autoencoder are fed into a GRU, which outputs a normality confidence score. For windows with low normality confidence (the lowest 30% under the default budget), the expensive model is additionally run, and its score is used instead.

Critical strengths
  • The idea of using a routing function to decide between a fast model and an expensive model to reduce inference cost is reasonable, although it is somewhat common and, as the related work section notes, has already been explored, albeit to a limited extent, in time series anomaly detection.

  • The experimental scope is broad (8 datasets, 14 baselines, 4 CDMs).

Critical weaknesses
  • It is unclear why CARE improves detection performance over the expensive model. The improvement seems to imply that the cheap model performs better than the expensive model on the bypassed windows, which is counterintuitive. Rather, the results seem to suggest that the cheap model alone performs about as well as the expensive model.

  • The gating function requires additional training. It weakens the claim that CARE is an efficient method.

  • The design of the cheap model itself lacks novelty.

  • Figure 4 appears inconsistent with the tables. For example, Figure 4(c) shows about 7.05 s for standalone TranAD on ECG, whereas Tables 1, 3 and 10 report 9.050 s. Please clarify.

Questions for authors
  • What do the authors consider to be the novelty of this paper?
  • Is CARE still efficient when the training time is also taken into account?
  • What causes the discrepancy in Figure 4?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.