Generalist Representation, Specialist Detection: TS-Router for Time-Series Anomaly Detection

Tian Lan; Yifei Gao; Yimeng Lu; Xuming An; Meng Wang; Yue Pan; Wenjun He; Chenghao Liu

cs.LGOct 1, 2026

Generalist Representation, Specialist Detection: TS-Router for Time-Series Anomaly Detection

Authors: Tian Lan, Yifei Gao, Yimeng Lu, Xuming An, Meng Wang, Yue Pan, Wenjun He, Chenghao Liu, +1 more

Organizations: Department of Industrial Engineering Tsinghua University · Huawei · Datadog AI Research Paris, France

Abstract

Time-series anomaly detection (TSAD) is difficult to generalize across datasets because heterogeneous temporal dynamics imply different notions of normality and favor different detection criteria. While time-series foundation models provide transferable representations, coupling them with a fixed anomaly-scoring mechanism can overlook this variation. This motivates a different perspective on foundation-model-based TSAD: using foundation models to coordinate specialized anomaly criteria rather than directly imposing a universal one. Based on this view, we propose \textbf{TS-Router}, a generalist-representation, specialist-detection framework that estimates the relative competence of heterogeneous anomaly detectors from pretrained temporal representations and selects suitable specialists for each target series. To avoid relying on specialist-performance labels from real tasks, we derive soft competence supervision from specialists' relative performance on labeled simulated tasks. At deployment, routing requires no target anomaly labels, and only the selected specialists are fitted unsupervisedly on the target series. We bound Top- $k$ set-competence regret under representation coverage and conditional competence stability. Across 16 real-world benchmarks and four complementary evaluation metrics, TS-Router achieves the best overall average rank. Controlled ablations with multiple frozen TSFM encoders further support the use of pretrained representations for competence estimation and adaptive specialist selection. The code is available at https://anonymous.4open.science/r/TS-Router-D8FF.

Figures & tables

Figure 1: Domain-wise VUS-PR of representative specialists relative to TimeRCD, normalized by TimeRCD performance.

Figure 2: Overview of the TS-Router framework.

Method	VUS-PR		Aff.-F1		F1 T		Std.-F1		Overall Rank	#Top-1	#Top-2
Method	Score $\uparrow$	Rank $\downarrow$	Score $\uparrow$	Rank $\downarrow$	Score $\uparrow$	Rank $\downarrow$	Score $\uparrow$	Rank $\downarrow$	Overall Rank	#Top-1	#Top-2
Foundation-guided specialist routing
TS-Router	49.59	2.44	86.87	2.44	47.37	3.00	45.34	2.75	2.66	26	39
Direct zero-shot models
TimeRCD	37.00	3.88	83.94	3.69	40.88	3.62	37.03	4.19	3.84	14	26
DADA	30.16	5.62	81.41	4.88	37.29	5.84	32.60	6.31	5.66	5	10

Table 1: Overall performance across 16 real-world TSAD benchmarks. Score denotes the mean performance across datasets, and Rank denotes the average dataset-wise rank among the 12 compared methods, with ties assigned average ranks. Overall Rank averages ranks across all four metrics. #Top-1 and #Top-2 count first-place and top-two finishes across the 64 dataset–metric combinations. Methods follow different deployment paradigms without using target anomaly labels for model fitting. Best results are in bold and second-best results are underlined .

Selection and fusion strategy
Variant	Detection performance				Routing quality
Variant	VUS-PR $\uparrow$	Aff.-F1 $\uparrow$	F1 T $\uparrow$	Std.-F1 $\uparrow$	NDCG@ $k$ $\uparrow$	Hit@ $k$ $\uparrow$
Best Fixed	39.68	84.34	41.01	36.43	0.607	0.207
Full Ensemble	49.54	85.55	46.22	43.48	–	–
Best Fixed Top-3	47.44	86.56	47.56	43.42	0.625	0.532
Router Top-1	49.00	86.80	43.47	41.41	–	–

Table 2: Routing effectiveness and representation ablations on eleven univariate benchmarks. Best Fixed denotes the hindsight-best single specialist shared across targets for each evaluation metric, whereas Oracle denotes hindsight per-target individual-specialist selection; these are single-specialist detection references, not Top-3 fusion oracles. Best Top-3 always fuses the hindsight-best set of three specialists shared across targets. Routing metrics use

k{=}1

for Best Fixed and Oracle and

k{=}3

otherwise; Router Top-1 shares the TS-Router ranking, so its routing metrics are not repeated. Representation variants differ only in the router representation, with the same router architecture, specialist pool, supervision, and Top-3 normalized mean fusion. The dashed line separates conventional feature representations from frozen TSFM encoders (Chronos, MOMENT, TimeRCD); TimeRCD corresponds to the default TS-Router configuration.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Name	Domain	#TS	Avg. Length	AR (%)
UCR	Misc.	228	67818.7	0.6
NAB	Mixed	28	5099.7	10.6
YAHOO	Web	259	1560.2	0.6
IOPS	Operations	17	72792.3	1.3
MGAB	Synthetic	9	97777.8	0.2
SED	Medical	3	23332.3	4.1

Appendix

Table 3: Statistics of the univariate benchmark datasets.

Name	Domain	#TS	Avg. Length	AR (%)
MSL	Space	16	3119.4	5.1
PSM	Sensor	1	217624.0	11.2
SMAP	Space	27	7855.9	2.9
SMD	Server	22	25466.4	3.8
SWaT	ICS	2	207457.5	12.7

Appendix

Table 4: Statistics of the multivariate benchmark datasets.

Specialist	Primary inductive bias
PCA	Low-dimensional subspace reconstruction; anomalies produce large reconstruction residuals.
Sub-PCA	Windowed PCA on subsequences; anomalies produce large reconstruction residuals relative to local temporal structure.
Sub-HBOS	Marginal density modeling through histogram-based statistics; low-density observations receive larger anomaly scores.
POLY	Smooth temporal structure modeled by polynomial fitting; deviations from the fitted temporal trend indicate anomalies.
Sub-KNN	Neighborhood-distance structure; observations far from their nearest neighbors are considered anomalous.
Sub-LOF	Relative local density; anomalies exhibit substantially lower local density than their surrounding neighborhoods.

Appendix

Table 5: Specialists and their primary anomaly-detection inductive biases.

Specialist	Hyperparameters
PCA	Period rank $1$ ; all principal components.
Sub-PCA	Period rank $1$ ; all principal components.
Sub-HBOS	Period rank $1$ ; $10$ histogram bins.
POLY	Period rank $1$ ; polynomial degree $4$ .
Sub-KNN	Period rank $2$ ; $50$ neighbors.
Sub-LOF	Period rank $2$ ; $30$ neighbors.

Appendix

Table 6: Specialist hyperparameters used for competence construction and target deployment.

Metric	Model	Univariate datasets											Multivariate datasets
Metric	Model	IOPS	MGAB	NAB	NEK	Power	SED	Stock	TODS	UCR	WSD	YAHOO	MSL	PSM	SMAP	SMD	SWaT
VUS-PR	TS-Router	41.49	43.09	51.97	76.92	22.40	65.42	73.18	71.65	39.60	56.91	77.32	20.86	19.55	38.12	52.24	42.75
	TimeRCD	20.23	1.05	24.32	27.88	21.25	80.75	77.28	93.46	23.09	21.77	84.41	20.45	18.69	22.68	37.03	17.58
	DADA	24.97	0.57	24.73	46.85	10.61	6.42	99.51	64.83	2.94	33.42	70.74	12.74	17.17	20.02	25.98	21.13
	Chronos	19.00	0.60	23.76	31.80	10.95	8.65	97.49	70.66	6.56	18.81	83.54	8.25	14.61	5.18	10.22	16.44
	MOMENT	37.35	0.56	45.38	67.74	10.50	4.31	76.97	56.45	6.17	55.26	30.81	9.32	16.48	8.97	15.96	14.90

Appendix

Table 7: Per-dataset performance across 16 real-world TSAD benchmarks. Best results are in bold and second-best results are underlined .

Metric	Representation	Univariate datasets											Avg.
Metric	Representation	IOPS	MGAB	NAB	NEK	Power	SED	Stock	TODS	UCR	WSD	YAHOO	Avg.
VUS-PR	TimeRCD	41.49	43.09	51.97	76.92	22.40	65.42	73.18	71.65	39.60	56.91	77.32	56.36
	Chronos	38.63	33.44	49.34	50.10	15.99	94.18	71.64	78.81	33.17	53.83	66.87	53.27
	MOMENT	40.73	33.44	46.83	33.93	20.56	94.18	87.24	76.35	38.09	57.55	63.87	53.89
	Catch22	44.58	9.46	42.75	38.31	22.27	94.18	76.57	84.05	34.11	48.86	54.39	49.96
	TSFresh	36.98	35.44	35.53	50.10	17.54	37.17	88.43	73.84	33.18	51.50	79.40	49.01

Appendix

Table 8: Representation ablation on eleven univariate TSB-AD benchmarks. All variants use the same MLP, 11-specialist pool, and Top-3 normalized mean fusion; only the router representation changes. TimeRCD remains frozen. Best results are in bold and second-best results are underlined .

Metric	Method	Univariate datasets											Avg.
Metric	Method	IOPS	MGAB	NAB	NEK	Power	SED	Stock	TODS	UCR	WSD	YAHOO	Avg.
VUS-PR	TS-Router	41.49	43.09	51.97	76.92	22.40	65.42	73.18	71.65	39.60	56.91	77.32	56.36
	PCA	23.21	0.60	45.77	78.45	10.49	3.77	80.92	54.01	13.40	16.95	21.15	31.70
	Sub-PCA	20.35	0.66	49.82	83.68	10.49	4.00	77.81	52.82	13.53	14.09	19.41	31.51
	Sub-HBOS	4.97	0.50	33.26	19.43	15.69	63.09	64.60	64.54	15.07	2.21	12.08	26.86
	POLY	27.51	0.70	46.91	55.64	9.30	7.88	79.01	57.39	15.80	35.24	27.89	33.02

Appendix

Table 9: Specialist complementarity on eleven univariate TSB-AD benchmarks. TS-Router uses Top-3 normalized mean fusion, while Oracle denotes hindsight per-target selection of the best individual specialist. Oracle is an individual-specialist reference rather than an upper bound on the fused TS-Router prediction. Best and second-best non-oracle results are in bold and underlined , respectively.

Figure 3: Effect of the number of selected specialists

k

with the 11-specialist pool. Each panel reports the unweighted mean over eleven univariate benchmarks; the panels show VUS-PR, F1 T , Standard-F1, and Affiliation-F1.

Figure 4: Effect of specialist-pool size with Top-3 selection. Each panel reports the unweighted mean over eleven univariate benchmarks; the panels show VUS-PR, F1 T , Standard-F1, and Affiliation-F1. The router is retrained for each pool size.

Routing head	VUS-PR	VUS-ROC	F1 T	Std.-F1	Aff.-F1
VUS-PR	56.36	88.30	49.62	47.73	88.11
Affiliation-F1	56.27	89.24	49.17	46.83	88.49
F1 T	58.24	89.29	53.03	49.96	89.17
Standard-F1	55.28	87.81	50.00	47.77	87.83

Appendix

Table 10: Top-3 routing with different competence-supervision heads. Values are dataset-averaged percentages over eleven univariate benchmarks. The pool size is 11 and the router is trained with the indicated head.

Metric	Selector	Univariate datasets											Avg.
Metric	Selector	IOPS	MGAB	NAB	NEK	Power	SED	Stock	TODS	UCR	WSD	YAHOO	Avg.
VUS-PR	TS-Router	41.49	43.09	51.97	76.92	22.40	65.42	73.18	71.65	39.60	56.91	77.32	56.36
	SATzilla	41.77	33.44	41.71	53.12	22.40	74.65	70.43	72.73	41.82	58.30	59.20	51.78
	ARGOSMART	44.02	9.46	44.73	31.94	22.36	11.94	82.21	70.71	34.24	54.88	53.37	41.81
	MetaOD	42.21	33.44	40.23	57.82	22.40	33.96	70.20	72.73	41.82	56.72	57.89	48.13
	ISAC	43.02	33.44	41.84	39.78	22.40	74.65	76.03	72.72	40.42	58.08	57.09	50.86

Appendix

Table 11: Algorithm-selection transfer on eleven univariate TSB-AD benchmarks. Classical selectors are trained on Syn-RCD using the common 11-specialist pool. TS-Router uses its fixed routing policy learned from the same synthetic competence supervision. All selectors use Top-

3

fusion of the predicted specialists. Best results are in bold and second-best results are underlined .

Metric	Method	Univariate datasets										Avg.
Metric	Method	IOPS	MGAB	NAB	NEK	SED	Stock	TODS	UCR	WSD	YAHOO	Avg.
VUS-PR	TS-Router	28.72	45.61	57.45	94.24	63.78	52.84	87.54	29.67	45.05	69.09	57.40
	SATzilla	20.45	48.52	43.74	76.40	77.57	90.36	73.87	38.32	35.01	45.81	55.00
	ARGOSMART	23.07	48.52	42.86	41.35	77.57	94.57	78.03	25.53	38.48	49.53	51.95
	MetaOD	19.58	48.52	39.00	76.40	77.57	90.36	77.94	39.62	35.01	55.13	55.91
	ISAC	18.96	42.10	42.93	36.70	36.09	52.00	64.30	31.10	33.26	19.34	37.68

Appendix

Table 12: In-distribution algorithm selection on the official TSB-AD Tuning subset. Evaluation uses the thirty genuinely univariate Tuning series from ten collections; Power does not appear in the Tuning split. Classical selectors are adapted from a small labeled support set across ten random trials, whereas TS-Router uses a fixed routing policy without support-set adaptation. Best results are in bold and second-best results are underlined .

Metric	Method	Held-out target dataset											Avg.
Metric	Method	IOPS	MGAB	NAB	NEK	Power	SED	Stock	TODS	UCR	WSD	YAHOO	Avg.
VUS-PR	TS-Router	41.49	43.09	51.97	76.92	22.40	65.42	73.18	71.65	39.60	56.91	77.32	56.36
	SATzilla	29.32	0.97	38.66	39.31	24.97	7.88	77.81	62.13	20.21	43.48	29.34	34.01
	ARGOSMART	27.68	24.70	32.40	52.31	9.30	24.70	76.93	57.39	20.83	44.15	36.26	36.97
	MetaOD	27.79	0.97	40.12	39.31	24.97	7.88	77.81	63.46	20.43	44.46	29.37	34.23
	ISAC	31.77	24.70	41.41	41.83	37.05	7.88	78.19	64.33	25.04	40.10	30.47	38.43

Appendix

Table 13: Leave-one-dataset-out transfer across eleven univariate benchmarks. Classical selectors are trained on the other ten real benchmark collections and deploy their top-ranked specialist (Top-1). TS-Router uses its fixed routing policy learned on Syn-RCD and default Top-3 normalized mean fusion, without real-data retraining of the router. All methods are evaluated on the entirely held-out target dataset. Best results are in bold and second-best results are underlined .

Explore similar work

ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection
May 31, 2026Uzair Khan, Luigi Capogrosso, Francesco Biondani +4Time-Series Anomaly Detection Time Series Foundation Models
University of Verona · Interdisciplinary Transformation University of Austria · ETH Zurich
VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection
Mar 27, 2026PengYu Chen, Shang Wan, Xiaohou Shi +3Time-Series Anomaly Detection Unsupervised Detection
School of Computer Science (National Pilot Software Engineering School), Beijing University of Posts and Telecommunications, Beijing 100876, China · China Telecom Research Institute Beijing, China · Department of Computer Science, Missouri University of Science and Technology, Rolla, MO 65409 USA
PaAno: Patch-Based Representation Learning for Time-Series Anomaly Detection
Feb 1, 2026Jinju Park, Seokho KangTime-Series Anomaly Detection Time Series
Department of Industrial Engineering, Sungkyunkwan University

May 31, 2026, cs.LGJ/K move · Enter open · S save

ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection

Uzair Khan, Luigi Capogrosso, Francesco Biondani, Michele Magno+3

University of Verona · Interdisciplinary Transformation University of Austria · ETH Zurich

Time-Series Anomaly Detection Time Series Foundation Models Time Series Chronos Bidirectional Long Short-Term Memory

Open paper AskPDF arXiv

Time series anomaly detection is a crucial task in various domains, including finance, healthcare, and industry. However, existing methods often struggle to generalize across different datasets, especially when anomalies are subtle or context-dependent. To solve this issue, we introduce ChronosAD, a novel architecture for anomaly detection that uses a time series foundation model as a feature extractor. Specifically, it employs a two-stage pipeline: first, it uses the foundation model to extract embeddings for each time series in a zero-shot manner. Then, a custom-developed Temporal Block, composed of Bidirectional Long Short-Term Memory (BiLSTM) and Multi-Head Attention, refines these embeddings to capture temporal dependencies and highlight salient patterns. Unlike previous approaches, our model requires minimal task-specific tuning and demonstrates robust generalization across a wide range of domains, including industrial, medical, cyber-physical, and automotive systems. Extensive experiments on 11 benchmarks show that ChronosAD outperforms existing methods by 4.72% in AUC and 6.60% in AP on average. The source code is available at https://github.com/intelligolabs/ChronosAD.

Abstract

Explore similar work

ChronosAD: Leveraging Time Series Foundation Models for Accurate Anomaly Detection

VAN-AD: Visual Masked Autoencoder with Normalizing Flow For Time Series Anomaly Detection

PaAno: Patch-Based Representation Learning for Time-Series Anomaly Detection