Pith. sign in

REVIEW 6 major objections 5 minor 2 cited by

Tool-augmented reasoning lifts F1 for time-series anomaly detection

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-02 23:24 UTC pith:A756ZO2D

load-bearing objection AnomaMind's architecture is a real contribution, but Table 2's F1 numbers don't match its own precision/recall columns, and the asymmetric training protocol leaves the central empirical claim unsupported. the 6 major comments →

arxiv 2602.13807 v2 pith:A756ZO2D submitted 2026-02-14 cs.LG

AnomaMind: Agentic Time Series Anomaly Detection with Tool-Augmented Reasoning

Xiaoyu Tao , Yuchong Wu , Mingyue Cheng , Ze Guo , Tian Gao This is my paper
classification cs.LG
keywords time series anomaly detectionagentic workflowtool-augmented reasoningreinforcement learninglarge language modelcoarse-to-fine detectiondiagnostic evidencegeneralization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that time series anomaly detection should be treated not as a one-shot scoring task but as an evidence-driven diagnostic process. To test this, it builds AnomaMind, an agentic framework that first localizes suspicious intervals, then gathers numerical and contextual evidence through modular tools, and finally refines its decision by self-reflection, with the core detection policy trained by reinforcement learning. The paper reports that this system outperforms statistical, deep-learning, foundation-model, and LLM baselines on four benchmarks, with the largest gains on datasets with complex anomaly types. If these results hold, they suggest that giving detectors the ability to adaptively prepare features and revise decisions is more valuable than any single static architecture.

Core claim

On the paper's own terms, the central discovery is that reformulating time series anomaly detection as a sequential decision-making process—rather than a discriminative prediction with fixed features—makes detection more accurate and more generalizable. AnomaMind combines a coarse-to-fine workflow with a toolkit of reusable analysis tools and a hybrid inference mechanism that separates flexible reasoning from task-specific policy learning. The paper claims this achieves consistently higher F1 scores than existing baselines across four diverse benchmarks, and interprets the result as evidence that tool-augmented, iterative reasoning is a viable alternative to model-centric anomaly detection.

What carries the argument

The load-bearing mechanism is a hybrid inference loop. A general-purpose large language model handles the flexible parts—visual interval localization, tool invocation, and self-reflection—while a separate detection policy is optimized by reinforcement learning with rule-based rewards that penalize unparsable outputs, align with F1-score, and suppress false positives. The workflow itself is four stages: coarse evidence acquisition (locating suspicious intervals via a vision-language model), adaptive evidence construction (invoking statistical, value-based, change-based, and region-level tools), reasoning-based detection (the RL-trained detector), and iterative refinement (a lenient evaluator

Load-bearing premise

The central claim rests on the assumption that the F1-alignment term in the reward and the extra training sequences drawn from the same benchmarks do not leak information about test labels; if they do, the reported superiority over baselines is an artifact of the evaluation protocol.

What would settle it

Retrain AnomaMind on one benchmark (or all four) with the F1-alignment reward term dropped (or computed on a held-out label set) and with training restricted to the exact anomaly-free prefix used for baselines. If the F1 advantage over the best baseline disappears or falls within noise, the paper's central claim that agentic reasoning itself improves detection is falsified; if the advantage persists, it is supported.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the reported gains are real, anomaly detection systems can move from static models to agentic pipelines that produce an interpretable audit trail of evidence, including which intervals were examined, which tools were invoked, and why the verdict passed or failed.
  • The hybrid design supplies a template for combining LLM reasoning with RL-trained decision modules in other diagnostic tasks, such as root-cause analysis or model monitoring, where a flexible planner can gather evidence and a specialized policy makes the final call.
  • The claim that tool-augmented reasoning improves generalization implies that evaluation protocols for time series anomaly detection should include cross-domain and concept-shift settings, not just in-distribution test splits.
  • The gains on the low-anomaly-rate, real-world datasets point to practical value in production monitoring, where false positives are expensive and the cost of an agentic workflow may be justified.
  • The ablations show that removing the RL-trained detector collapses performance, suggesting the framework's success depends more on the task-specific decision policy than on the language model's generic reasoning.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's reward includes an explicit F1-alignment term, and its training protocol draws extra sequences from the same benchmarks, while baselines train only on an anomaly-free prefix. Unless those labels and sequences are provably disjoint from the evaluation labels, the magnitude of the reported gains should be read as an upper bound; a reader who wants to know how much the agentic design itse
  • A natural extension is to test AnomaMind with the reward F1 term removed and with training restricted to the same prefix as baselines, on the same four datasets, to see whether the agentic workflow alone—without any in-distribution training advantage—still beats the best baseline. This concrete experiment would separate the framework's contribution from its training protocol.
  • If the approach is correct, one would expect it to transfer to multivariate and streaming settings where contextual evidence is richer; the tools could be extended to correlation and causality operators, and the refinement step could become a cost-sensitive stopping rule, trading token budget against detection confidence.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

6 major / 5 minor

Summary. The paper proposes AnomaMind, an agentic framework for time series anomaly detection that reformulates detection as a sequential decision-making process. The proposed workflow is coarse-to-fine: a vision-language locator identifies candidate intervals, tool-assisted actors gather numerical and contextual evidence, a detector issues fine-grained decisions, and an evaluator refines them. The authors introduce a hybrid inference mechanism in which a general-purpose LLM handles reasoning and tool orchestration while a detection-specific policy is trained by reinforcement learning with rule-based rewards, including an F1-alignment term. The paper claims consistent F1 improvements over ten statistical, deep-learning, foundation-model, and LLM baselines across four benchmarks, and supports this with ablations over tools, RL rewards, backbone sizes, and model choices. The code is released.

Significance. If the empirical claims were reliable, the paper would make a timely contribution by connecting agentic LLM workflows with time series anomaly detection and by separating flexible reasoning from task-specific decision learning. The code release, the explicit tool design, and the breadth of ablations are strengths. However, the quantitative evidence is currently undermined by internal inconsistencies in the main results table, an unexplained mismatch between the main configuration and the ablation configurations, asymmetric training data for AnomaMind versus baselines, and the use of F1 both as an RL reward component and as the primary evaluation metric without demonstrated label disjointness. Until these issues are resolved, the central claim of consistent improvement is not supported.

major comments (6)
  1. [Table 2 / §D.3.1] Under the F1 definition given in §D.3.1 (harmonic mean of Precision and Recall), many rows in Table 2 are internally inconsistent. For example, YAHOO LSTMAD reports P=0.394, R=0.449 but F1=0.360, whereas 2PR/(P+R)=0.420; YAHOO TransAD reports P=0.007, R=0.015 but F1=0.082, whereas the harmonic mean is 0.010; and the AnomaMind row reports P=0.841, R=0.837 with F1=0.821, not 0.839. No note in the paper explains an alternative F1 variant. Since Table 2 is the only support for the claim of consistent improvement, this inconsistency is load-bearing.
  2. [§5.1.3, Tables 2–6] The main result is not reproducible from any configuration described in the paper. Section 5.1.3 states that Qwen3-8B is the fine-grained reasoning backbone and grok-4 is used for all other modules, but Table 2 (YAHOO F1=0.821, Best-F1=0.835) does not match Table 4's Qwen3-8B row (0.793/0.821), Table 6's Grok-4 row (0.768/0.778), or the 'Ours' rows of Tables 3 and 5 (0.713/0.715). Please identify the exact configuration, reconcile the numbers, or rerun the experiment.
  3. [§D.2] The comparison in Table 2 uses asymmetric data protocols. Baselines train on the anomaly-free initial segment of the test sequence, while AnomaMind trains on 'additional distinct sequences from the same dataset' (App. D.2). This gives AnomaMind access to more training data and potentially to information about the test distribution. The claimed gains may therefore reflect data advantage rather than the proposed mechanism. The authors should retrain baselines under the same protocol or evaluate AnomaMind under the baseline protocol, and report both.
  4. [§4.4, Figure 1, §D.3.2] The RL reward in Figure 1 includes an explicit 'F1score Reward' term, and the paper's primary metric is F1 (Table 2). To rule out circularity, the authors must demonstrate that the labels used for the F1 reward are disjoint from the evaluation labels; this is never stated. Moreover, Best-F1 for AnomaMind is computed by sweeping the confidence threshold on test data (§D.3.2), which is an optimistic selection on the evaluation set. The paper should clarify the train/test split for reward computation and report a fixed-threshold F1 as well.
  5. [Abstract, §5] The abstract claims 'extensive experiments under both in-domain and cross-domain settings', but Section 5 contains only in-domain experiments on four benchmarks (Table 1) and no cross-domain protocol or results. This claim is not supported. The authors must add cross-domain experiments (e.g., train on one dataset, test on another) or remove the claim from the abstract.
  6. [§5.1.3, Table 2] All tables and figures report single-run values, despite the stochastic nature of LLM sampling and RL optimization. The performance differences between AnomaMind and strong baselines are sometimes only a few points, so without repeated-seed results with means and standard deviations the improvements cannot be distinguished from run-to-run variability.
minor comments (5)
  1. [§5.2] The sentence ending '...conventional model-centric paradigms' is repeated twice in the same paragraph; delete the duplicate.
  2. [Figure 4] The boxes in the case-study figure have clipped/overlapping text (e.g., 'Diff Z-Score: Detected a dense outlier cluster' and 'Verdict: Pass'), making the workflow hard to follow.
  3. [Figure 6] No legend identifies which curve corresponds to which dataset or backbone; the caption says 'different datasets and backbone models' but the reader cannot tell them apart.
  4. [Appendix C.3] The phrase 'lenientevaluation stance' is a typo ('lenient evaluation').
  5. [Figure 7 / §B] The baseline name is inconsistently written as 'TranAD' in Figure 7 and 'TransAD' in the text; standardize.

Circularity Check

0 steps flagged

No significant circularity: the F1-aligned RL reward is a standard training objective computed on training sequences, not a construction of the reported test F1.

full rationale

The claimed derivation chain is an empirical system built on a coarse-to-fine agentic workflow, tool invocation, and an RL-trained detection policy. The reader's 6.0 suspicion focuses on the Fig. 1 'F1score Reward' and the paper's use of F1 as the primary metric. But the paper describes training on separate data: Table 1 gives distinct Train Points and Test Points, and App. D.2 states 'for our proposed method, we incorporate additional distinct sequences from the same dataset as training samples to enhance generalization.' Nothing in the text indicates that the labels used to compute the F1 reward during RL are the same labels used to compute the reported test F1. Optimizing a training reward that aligns with the final evaluation metric is standard supervised/RL practice, not circularity. The self-citation [11] is used to motivate the agentic framing, but the same claim is supported by independent citations and the central empirical claim rests on comparisons against 10 external baselines, not on [11]. Separate correctness concerns exist: Table 2's F1 values do not always match the paper's own F1 definition in App. D.3.1 (e.g., LSTMAD on YAHOO reports P=0.394, R=0.449, F1=0.360, while the harmonic mean is 0.420), and the AnomaMind row is not reproducible from the stated configurations in Tables 3, 4, and 6. These are internal-consistency and reproducibility issues, not circularity. No load-bearing step in the claimed derivation reduces, by definition or by self-citation, to its own inputs.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework introduces no new physical entities, but its performance claim rests on several unverified modeling assumptions and unreported hyperparameters. The most consequential are the visual saliency of anomalies and the legitimacy of using F1-aligned rewards without a demonstrated label split.

free parameters (4)
  • Best-F1 confidence threshold = swept per dataset
    Best-F1 is obtained by sweeping thresholds over confidence scores (§D.3.2), selecting the most favorable operating point post hoc.
  • sample length and step size = 100 / 100
    Chosen in §D.2; determines context window and segmentation and may affect localization and evidence quality.
  • RL reward weights = not reported
    The relative weights of parsing, F1-alignment, and false-positive rewards are not given, so the training objective is under-specified.
  • temperature = not stated (sensitivity range 0.0–1.0)
    Default sampling temperature is never reported; §C.3 shows performance varies with temperature.
axioms (4)
  • domain assumption Anomalies are visually detectable in normalized time-series plots.
    Interval localization uses a VLM on visual representations (§4.2); if anomalies are not visually salient, the coarse-to-fine workflow starts from unreliable candidates.
  • domain assumption Statistical and structural operators (z-scores, pattern features) provide sufficient evidence to confirm or reject anomalies.
    The detector's final decision is grounded in tool outputs (§4.3, §4.4); if these heuristics cannot separate anomalies from benign fluctuations, the framework has no independent signal.
  • domain assumption RL reward computed from F1 alignment generalizes from training labels to the test distribution.
    Reward design (§4.4, Table 5) assumes optimizing F1-style reward on available labels improves held-out F1; no cross-domain experiment validates this transfer.
  • domain assumption The general-purpose LLM/VLM executes tool calls and parses outputs reliably.
    The entire workflow depends on the LLM following structured plans and prompts (Appendix D.4); failure would break evidence construction.

pith-pipeline@v1.3.0-alltime-deepseek · 19366 in / 13355 out tokens · 114181 ms · 2026-08-02T23:24:22.260306+00:00 · methodology

0 comments
read the original abstract

Time series anomaly detection is critical in many real-world applications, where effective solutions must localize anomalous regions and support reliable decision-making under complex settings. However, most existing methods frame anomaly detection as a purely discriminative prediction task with fixed feature representations, rather than an evidence-driven diagnostic process. As a result, they often struggle when anomalies exhibit strong context dependence, diverse patterns, or domain shifts across datasets. To address these challenges, we propose AnomaMind, an agentic time series anomaly detection framework that reformulates anomaly detection as a sequential decision-making process. AnomaMind operates through a coarse-to-fine workflow that first localizes suspicious intervals, then constructs diagnostic evidence through tool interaction, and finally refines anomaly decisions through self-reflection. The workflow is supported by a toolkit box that combines knowledge memory and numerical diagnostics: visual anomaly patterns mined from training data and domain knowledge provide contextual guidance, while statistical, value-based, change-based, and region-level operators provide measurable evidence for verification. AnomaMind further adopts a hybrid inference mechanism in which general-purpose models handle flexible reasoning, tool invocation, and refinement, while a detection-specific policy is optimized with rule-based rewards for parsable outputs, F1-score alignment, and false-positive control. Extensive experiments under both in-domain and cross-domain settings demonstrate that AnomaMind consistently improves anomaly detection performance and enhances generalization across heterogeneous anomaly patterns, validating the effectiveness of tool-augmented reasoning for anomaly detection. The code is available at https://github.com/Xiaoyu-Tao/AnomaMind-TS.

Figures

Figures reproduced from arXiv: 2602.13807 by Mingyue Cheng, Tian Gao, Xiaoyu Tao, Yuchong Wu, Ze Guo.

Figure 1
Figure 1. Figure 1: Overview of AnomaMind, an agentic TSAD framework that adopts a structured coarse-to-fine workflow built upon a [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Performance comparison with and without the eval [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Performance comparison with and without rein [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: An illustrative example of a structured coarse-to-fine workflow for agentic time series anomaly detection. [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of different datasets and backbone models [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 5
Figure 5. Figure 5: Evaluation of distinct specialized module variants [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of detection results demonstrating the superior precision of AnomaMind compared to leading baselines. [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Hyperparameter sensitivity analysis regarding sam [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. CastFlow: Learning Role-Specialized Agentic Workflows for Time Series Forecasting

    cs.LG 2026-04 unverdicted novelty 7.0

    CastFlow introduces a role-specialized agentic workflow with memory retrieval and multi-view toolkit for iterative ensemble time series forecasting, using two-stage SFT+RLVR training on a domain-specific LLM to outper...

  2. Failures Reveal What Metrics Miss: An Evidence-Driven Agent for Recursive Refinement of ECG Classifiers

    cs.AI 2026-07 conditional novelty 6.5

    An evidence-grounded LLM-as-Designer agent refines ECG classifiers from failure cases and deterministic measurements, freezing a stronger deployable model with ~10% relative macro-F1 gains.

Reference graph

Works this paper leans on

50 extracted references · 11 linked inside Pith · cited by 2 Pith papers

  1. [1]

    Sarah Alnegheimish, Linh Nguyen, Laure Berti-Equille, and Kalyan Veeramacha- neni. 2024. Large language models can be zero-shot anomaly detectors for time series?arXiv preprint arXiv:2405.14755(2024)

    Pith/arXiv arXiv 2024

  2. [2]

    Abdul Fatir Ansari, Lorenzo Stella, Caner Turkmen, Xiyuan Zhang, Pedro Mercado, Huibin Shen, Oleksandr Shchur, Syama Sundar Rangapuram, Sebas- tian Pineda Arango, Shubham Kapoor, et al. 2024. Chronos: Learning the language of time series.arXiv preprint arXiv:2403.07815(2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Julien Audibert, Pietro Michiardi, Frédéric Guyard, Sébastien Marti, and Maria A Zuluaga. 2020. Usad: Unsupervised anomaly detection on multivariate time series. InProceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining. 3395–3404

    2020

  4. [4]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. 2023. Qwen technical report.arXiv preprint arXiv:2309.16609(2023)

    Pith/arXiv arXiv 2023

  5. [5]

    Adya Bhat, Abhay Kumar, Akash Chekodu, BS Chandana, and SS Shylaja. 2025. Enhanced Anomaly Detection in Time-Series Data: A Comparative Study of Univariate Approach with Transformer and LLM Methods. InInternational Con- ference on Information and Communication Technology for Intelligent Systems. Springer, 261–271

    2025

  6. [6]

    Markus M Breunig, Hans-Peter Kriegel, Raymond T Ng, and Jörg Sander. 2000. LOF: identifying density-based local outliers. InProceedings of the 2000 ACM SIGMOD international conference on Management of data. 93–104

    2000

  7. [7]

    Varun Chandola, Arindam Banerjee, and Vipin Kumar. 2009. Anomaly detection: A survey.ACM computing surveys (CSUR)41, 3 (2009), 1–58

    2009

  8. [8]

    Akash Chandrayan, ZIDI Amir, Matthew Reimherr, Anis Mjirda, and Abhinav Pradhan. [n. d.]. LEAD-Framework for efficient time-series anomaly detection on large scale data using LLMs. In1st ICML Workshop on Foundation Models for Structured Data

  9. [9]

    Sucheta Chauhan and Lovekesh Vig. 2015. Anomaly detection in ECG time signals via deep long short-term memory networks. In2015 IEEE international conference on data science and advanced analytics (DSAA). IEEE, 1–7

    2015

  10. [10]

    Mingyue Cheng, Jie Ouyang, Shuo Yu, Ruiran Yan, Yucong Luo, Zirui Liu, Daoyu Wang, Qi Liu, and Enhong Chen. 2025. Agent-r1: Training powerful llm agents with end-to-end reinforcement learning.arXiv preprint arXiv:2511.14460(2025)

    Pith/arXiv arXiv 2025

  11. [11]

    Mingyue Cheng, Xiaoyu Tao, Qi Liu, Ze Guo, and Enhong Chen. 2026. Posi- tion: Beyond Model-Centric Prediction–Agentic Time Series Forecasting.arXiv preprint arXiv:2602.01776(2026)

    arXiv 2026

  12. [12]

    Abhimanyu Das, Weihao Kong, Rajat Sen, and Yichen Zhou. 2024. A decoder- only foundation model for time-series forecasting. InForty-first International Conference on Machine Learning

    2024

  13. [13]

    Manqing Dong, Hao Huang, and Longbing Cao. 2024. Can LLMs Serve As Time Series Anomaly Detectors?arXiv preprint arXiv:2408.03475(2024)

    Pith/arXiv arXiv 2024

  14. [14]

    Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. 2025. MOMENT: A Family of Open Time-series Foundation Models. InForty-first International Conference on Machine Learning

    2025

  15. [15]

    Yile Gu, Yifan Xiong, Jonathan Mace, Yuting Jiang, Yigong Hu, Baris Kasikci, and Peng Cheng. 2025. Argos: Agentic time-series anomaly detection with autonomous rule generation via large language models.arXiv preprint arXiv:2501.14170(2025)

    Pith/arXiv arXiv 2025

  16. [16]

    Zhaopeng Gu, Bingke Zhu, Guibo Zhu, Yingying Chen, Ming Tang, and Jinqiao Wang. 2024. Anomalygpt: Detecting industrial anomalies using large vision- language models. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 1932–1940

    2024

  17. [17]

    Manish Gupta, Jing Gao, Charu C Aggarwal, and Jiawei Han. 2013. Outlier detection for temporal data: A survey.IEEE Transactions on Knowledge and data Engineering26, 9 (2013), 2250–2267

    2013

  18. [18]

    2018.Forecasting: principles and practice

    Rob J Hyndman and George Athanasopoulos. 2018.Forecasting: principles and practice. OTexts

    2018

  19. [19]

    Chuang Jiang, Mingyue Cheng, Xiaoyu Tao, Qingyang Mao, Jie Ouyang, and Qi Liu. 2025. Tablemind: An autonomous programmatic agent for tool-augmented table reasoning.arXiv preprint arXiv:2509.06278(2025)

    Pith/arXiv arXiv 2025

  20. [20]

    Yushan Jiang, Zijie Pan, Xikun Zhang, Sahil Garg, Anderson Schneider, Yuriy Nevmyvaka, and Dongjin Song. 2024. Empowering time series analysis with large language models: a survey. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. 8095–8103

    2024

  21. [21]

    Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, et al. 2024. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. InThe Twelfth International Conference on Learning Representations

    2024

  22. [22]

    Dongmin Kim, Sunghyun Park, and Jaegul Choo. 2024. When model meets new normals: Test-time adaptation for unsupervised time-series anomaly detection. In Proceedings of the AAAI conference on artificial intelligence, Vol. 38. 13113–13121

    2024

  23. [23]

    Kwei-Herng Lai, Daochen Zha, Guanchu Wang, Junjie Xu, Yue Zhao, Devesh Kumar, Yile Chen, Purav Zumkhawaka, Minyang Wan, Diego Martinez, et al

  24. [24]

    Nikolay Laptev, Saeed Amizadeh, and Ian Flint. 2015. Generic and scalable framework for automated time-series anomaly detection. InProceedings of the 21th ACM SIGKDD international conference on knowledge discovery and data mining. 1939–1947

    2015

  25. [25]

    Zhuofeng Li, Haoxiang Zhang, Seungju Han, Sheng Liu, Jianwen Xie, Yu Zhang, Yejin Choi, James Zou, and Pan Lu. 2025. In-the-flow agentic system optimization for effective planning and tool use.arXiv preprint arXiv:2510.05592(2025)

    Pith/arXiv arXiv 2025

  26. [26]

    Chen Liu, Shibo He, Qihang Zhou, Shizhong Li, and Wenchao Meng. 2024. Large language model guided knowledge distillation for time series anomaly detection. InProceedings of the Thirty-Third International Joint Conference on Artificial Intelligence. 2162–2170

    2024

  27. [27]

    Jun Liu, Chaoyun Zhang, Jiaxu Qian, Minghua Ma, Si Qin, Chetan Bansal, Qingwei Lin, Saravan Rajmohan, and Dongmei Zhang. 2025. Large language models can deliver accurate and interpretable time series anomaly detection. InProceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

    2025

  28. [28]

    Yang Liu, Yixing Luo, Xiaofeng Li, Xiaogang Dong, Bin Gu, and Zhi Jin. 2026. Eval- uating Large Language Models for Time Series Anomaly Detection in Aerospace Software.arXiv preprint arXiv:2601.12448(2026)

    arXiv 2026

  29. [29]

    Pankaj Malhotra, Lovekesh Vig, Gautam Shroff, Puneet Agarwal, et al. 2015. Long short term memory networks for anomaly detection in time series. InProceedings, Vol. 89. 94

    2015

  30. [30]

    Mohsin Munir, Shoaib Ahmed Siddiqui, Andreas Dengel, and Sheraz Ahmed

  31. [31]

    Junwoo Park, Kyudan Jung, Dohyun Lee, Hyuck Lee, Daehoon Gwak, ChaeHun Park, Jaegul Choo, and Jaewoong Cho. [n. d.]. Delving into Large Language Models for Effective Time-Series Anomaly Detection. InThe Thirty-ninth Annual Conference on Neural Information Processing Systems

  32. [32]

    Marco AF Pimentel, David A Clifton, Lei Clifton, and Lionel Tarassenko. 2014. A review of novelty detection.Signal processing99 (2014), 215–249

    2014

  33. [33]

    Hansheng Ren, Bixiong Xu, Yaoliang Wang, Qi Yi, et al . 2019. Time-Series Anomaly Detection Service at Microsoft. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. 3009–3017. doi:10.1145/3292500.3330680

    arXiv 2019

  34. [34]

    Stuart W Roberts. 2000. Control chart tests based on geometric moving averages. Technometrics42, 1 (2000), 97–101

    2000

  35. [35]

    2003.Robust regression and outlier detection

    Peter J Rousseeuw and Annick M Leroy. 2003.Robust regression and outlier detection. John wiley & sons

    2003

  36. [36]

    Alicia Russell-Gilbert, Alexander Sommers, Andrew Thompson, Logan Cummins, Sudip Mittal, Shahram Rahimi, Maria Seale, Joseph Jaboure, Thomas Arnold, and Joshua Church. 2024. Aad-llm: Adaptive anomaly detection using large language models. In2024 IEEE International Conference on Big Data (BigData). IEEE, 4194–4203

    2024

  37. [37]

    Sebastian Schmidl, Phillip Wenig, and Thorsten Papenbrock. 2022. Anomaly detection in time series: a comprehensive evaluation.Proceedings of the VLDB Endowment15, 9 (2022), 1779–1797

    2022

  38. [38]

    Bernhard Schölkopf, John C Platt, John Shawe-Taylor, Alex J Smola, and Robert C Williamson. 2001. Estimating the support of a high-dimensional distribution. Neural computation13, 7 (2001), 1443–1471

    2001

  39. [39]

    Ya Su, Youjian Zhao, Chenhao Niu, Rong Liu, Wei Sun, and Dan Pei. 2019. Robust anomaly detection for multivariate time series through stochastic recurrent neural network. InProceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining. 2828–2837

    2019

  40. [40]

    Shreshth Tuli, Giuliano Casale, and Nicholas R Jennings. 2022. Tranad: Deep transformer networks for anomaly detection in multivariate time series data. arXiv preprint arXiv:2201.07284(2022)

    Pith/arXiv arXiv 2022

  41. [41]

    Xingjian Wu, Junkai Lu, Zhengyu Li, Xiangfei Qiu, Jilin Hu, Chenjuan Guo, Christian S Jensen, and Bin Yang. 2026. TimeART: Towards Agentic Time Series Reasoning via Tool-Augmentation.arXiv preprint arXiv:2601.13653(2026)

    arXiv 2026

  42. [42]

    Haowen Xu, Wenxiao Chen, Nengwen Zhao, Zeyan Li, Jiahao Bu, Zhihan Li, Ying Liu, Youjian Zhao, Dan Pei, Yang Feng, et al. 2018. Unsupervised anomaly detection via variational auto-encoder for seasonal kpis in web applications. In Proceedings of the 2018 world wide web conference. 187–196

    2018

  43. [43]

    Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. 2022. Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy. In International Conference on Learning Representations

    2022

  44. [44]

    Zahra Zamanzadeh Darban, Geoffrey I Webb, Shirui Pan, Charu Aggarwal, and Mahsa Salehi. 2024. Deep learning for time series anomaly detection: A survey. Comput. Surveys57, 1 (2024), 1–42

    2024

  45. [45]

    Shenglin Zhang, Zhenyu Zhong, Dongwen Li, Qiliang Fan, Yongqian Sun, Man Zhu, Yuzhi Zhang, Dan Pei, Jiyan Sun, Yinlong Liu, et al . 2022. Efficient kpi anomaly detection through transfer learning for large-scale web services.IEEE Journal on Selected Areas in Communications40, 8 (2022), 2440–2455

    2022

  46. [46]

    Xiaohan Zhang, Tian Gao, Mingyue Cheng, Bokai Pan, Ze Guo, Yaguo Liu, and Xiaoyu Tao. 2025. AlphaCast: A Human Wisdom-LLM Intelligence Co- Reasoning Framework for Interactive Time Series Forecasting.arXiv preprint arXiv:2511.08947(2025). AnomaMind: Agentic Time Series Anomaly Detection with Tool-Augmented Reasoning Conference acronym ’XX, June 03–05, 2018...

    Pith/arXiv arXiv 2025

  47. [47]

    Tian Zhou, Peisong Niu, Liang Sun, Rong Jin, et al . 2023. One fits all: Power general time series analysis by pretrained lm.Advances in neural information processing systems36 (2023), 43322–43355

    2023

  48. [48]

    interval

    Deyao Zhu, Jun Chen, Xiaoqian Shen, Xiang Li, and Mohamed Elhoseiny. 2023. Minigpt-4: Enhancing vision-language understanding with advanced large lan- guage models.arXiv preprint arXiv:2304.10592(2023). A Dataset Details To rigorously evaluate the effectiveness of our method, we selected four widely used benchmarks in the field of time series anomaly dete...

    Pith/arXiv arXiv 2023

  49. [2018]

    DeepAnT: A deep learning approach for unsupervised anomaly detection in time series.Ieee Access7 (2018), 1991–2005

    2018

  50. [2021]

    InProceedings of the aaai conference on artificial intelligence, Vol

    Tods: An automated time series outlier detection system. InProceedings of the aaai conference on artificial intelligence, Vol. 35. 16060–16062