Research map / Time series / Anomaly detection

STAR: Boosting Time Series Foundation Models for Anomaly Detection Through State-Aware Adapter

NeurIPS 2026 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes STAR, a plug-and-play state-aware adapter that lets time series foundation models (DADA, UniTS, Moment, Timer) use discrete state variables for multivariate anomaly detection during fine-tuning.

Problem

Time series foundation models for multivariate anomaly detection treat discrete state variables (e.g., valve on/off, day of week) like numerical variables, which fails to exploit their conditioning role and can even degrade detection when state variables are added.

Approach

An Identity-guided State Encoder that routes variable and state identities over a small learnable State Memory; a Conditional Bottleneck Adapter that generates low-rank adaptation parameters (from SVD-decomposed frozen weights) conditioned on the current state; and a Numeral-State Matching contrastive loss that detects state-level anomalies.

Key results

• On Genesis and NYC, STAR improves VUS-ROC for backbones by 8.73% and 9.17% respectively (as stated by the authors).
• Compared with standard fine-tuning, STAR improves VUS-ROC on task-specific (DADA) and task-general TSFMs by 6.93% and 6.07% on average.
• Ablation on MSL with DADA: full STAR raises V-R from 0.746 to 0.807 and V-P from 0.284 to 0.331.

Surprising findings

• Adding state variables to TSFMs through unified modeling often fails to improve or degrades detection performance.
• LoRA fine-tuning gives inconsistent and often marginal changes compared to standard fine-tuning.

Limitations

• Gains are modest on some datasets (e.g., SWaT absolute V-R values remain low), and improvements vary across backbones.
• Pseudo-covariate experiments on datasets without state variables show performance decreases in most cases, per the authors.

Experimental setup

Datasets: MSL SMAP SWaT Genesis NYC PSM PUMP ECG

Benchmarks: MSL SMAP SWaT Genesis NYC PSM PUMP ECG

Metrics: AUC-ROC VUS-ROC VUS-PR AUC-PR R-AUC-ROC R-AUC-PR Range-F1 Affiliated-F1 Accuracy

Baselines: DADA UniTS Moment Timer LoRA

Builds on: LoRA Mixture of Experts DADA UniTS Moment Timer

Closest papers by meaning

• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.87)
• Retrieval Over Training: Similarity search-based Model Selection for Time Series Anomaly Detection (NeurIPS 2026 · similarity 0.86)
• Structured Temporal Causality for Interpretable Multivariate Time Series Anomaly Detection (NeurIPS 2025 · similarity 0.85)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.85)
• Towards a General Time Series Anomaly Detector with Adaptive Bottlenecks and Dual Adversarial Decoders (ICLR 2025 · similarity 0.85)
• When Model Meets New Normals: Test-Time Adaptation for Unsupervised Time-Series Anomaly Detection (AAAI 2024 · similarity 0.85)