WorksComputer Science
Evaluating Language Model Features for Zero-Shot Time-Series Anomaly Detection Under Isolation Assumption
Abstract
Time-Series Anomaly Detection (TSAD) is a fundamentally challenging task: it is unsupervised, requires high local precision capabilities, and suffers from limited labelled data. Recent work has explored adapting Large Language Models (LLMs) to time-series via cross-modal transfer. However, existing approaches are poorly suited to Multivariate TSAD (MTSAD), as tokenisation often obscures fine-grained temporal and variables structures, and most methods optimise auxiliary objectives such as forecasting or reconstruction. We posit that these surrogate objectives become unreliable as model capacity increases. Highly expressive models risk reconstructing or forecasting anomalies, undermining the foundations of both strategies. Instead, we propose to use LLMs purely as feature extractors and directly analyse their induced feature spaces. In this work, investigate the ability of LLMs to provide useful representations for MTSADs in a zero-shot setting through classification. To this end, we study the separability of the extracted features, under the assumption that anomalous features should be isolated from normal ones. We analyse their temporal and contextual discriminative ability using simple step-wise tokenisation schemes, with one time-step per token, which preserve fine-grained local information. Their usability for MTSAD is assessed using an unsupervised Isolation Forest, enabling a direct evaluation of the resulting representation quality. We further investigate whether the observed performance can be attributed to the language-related pre-training of the LLMs or arises independently of it. Our results provide empirical insight into the ability of LLMs to represent multivariate time-series for anomaly detection and highlight broader applications of cross-modal foundation models.
Introduction
Time-series analysis is inherently challenging due to the high heterogeneity and domain dependence of available datasets, compounded by their limited sizes. These constraints hinder the development of general-purpose Time-Series Foundation Models (TSFMs) capable of addressing core tasks such as classification, forecasting, and anomaly detection.
Recent work has explored the use of Large Language Models (LLMs) for time-series processing [48, 19]. Trained on large-scale textual corpora, LLMs possess strong sequential modelling capabilities and have demonstrated notable cross-modal adaptability [25]. Leveraging pre-trained LLMs as processors for time-series data is therefore an appealing direction, benefiting from the maturity and scale of language modelling research. A central challenge lies in the mismatch between time-series data and textual representations. Unlike natural language, time-series consist of continuous numerical values with precise temporal and inter-variable relationships. Prior work has identified several directions to address this mismatch, including learned projections, model adaptation, and specialised tokenisers [35]; however, only a few strategies allow the use of LLMs for time-series analysis in a zero-shot manner. The popular direct tokenisation of Gruver et al. [13] converts numbers to text characters but compromises numerical precision and can fragment values across multiple tokens, disrupting temporal dependencies. These issues are exacerbated in multivariate settings, where serialisation destroys the multivariate structure, hinders dependencies between variables, and drastically inflates token count [6].
A deeper issue lies in how existing approaches prioritise general representation learning, overlooking the specific demands of Time-Series Anomaly Detection (TSAD). In particular, TSAD poses additional difficulties, as it requires high local sensitivity and is typically formulated in an unsupervised setting due to the absence of anomalies in the training data [7, 28]. Common processing strategies, particularly tokenisation schemes such as patching or quantisation, can obscure fine-grained temporal patterns and attenuate anomalies. Furthermore, most methods rely on auxiliary objectives, notably forecasting or reconstruction; while effective for broad evaluation, these objectives are not aligned with anomaly detection: sufficiently expressive models risk reconstructing or forecasting anomalous patterns as faithfully as normal data, thereby suppressing residual scores and leaving anomalies undetected by design [11, 46]. Consequently, higher model capacity and reconstruction fidelity do not guarantee superior anomaly detection; on the contrary, excessive reconstruction capability can actively degrade performance by absorbing anomalous signals.
To circumvent these limitations, we consider anomaly detection through the lens of One-Class Classification (OCC) [34], where normal patterns define the data distribution and anomalies correspond to deviations from it. This perspective naturally favours approaches that operate directly on feature representations to estimate abnormality [16]. In this context, we investigate features extracted from pre-trained foundation models such as LLMs: as these models continue to scale, assessing their ability to transfer to new modalities in a zero-shot setting becomes increasingly relevant.
In this direction, we propose to evaluate the extent to which LLM-induced representations isolate anomalous from normal patterns. We therefore investigate simple, zero-shot-compatible tokenisation strategies, with an emphasis on step-wise tokenisation where each time-step is mapped to a single token. This design limits the effect of local information erasure, making it particularly suitable for detecting point anomalies while not compromising the detection of other anomaly types, and allowing for a focus on the temporal processing of LLMs.
Our contributions are summarised as follows:
We discuss the limitations of current foundation model approaches applied to Multivariate Time-Series Anomaly Detection (MTSAD) and propose investigating language model features as an alternative approach, treating anomaly detection as a central objective.
We propose a unified evaluation protocol designed for imbalanced, unsupervised anomaly detection settings, based on an isolation prior for anomalies.
We provide an extensive empirical study across multiple LLM architectures, tokenisation methods, datasets and context sizes, offering insight into the zero-shot applicability of LLMs for MTSAD. We further investigate whether language knowledge is leveraged through trivial tokenisation strategies.
Related Work
The emergence of TSFMs marks a significant shift from task-specific architectures toward unified frameworks capable of addressing core evaluation tasks such as classification, short- and long-term forecasting, and anomaly detection [20, 29]. Recent advancements, exemplified by UniTS [10], demonstrate the feasibility of training models across these diverse tasks simultaneously, allowing for fine-tuning on anomaly detection datasets or direct zero-shot application to novel tasks. However, many current TSFMs remain constrained by a focus on proxy objectives, such as reconstruction or forecasting which are typically limited approaches in anomaly detection [15].
Parallel to this, the sequential nature of time-series data has led to the adoption of LLMs as powerful backbones for temporal analysis [48]. This direction is driven by the inherent multi-modal capabilities of LLMs and their proficiency in processing sequential dependencies [19]. Most research in this area focuses on forecasting-based objectives or fine-tuning strategies designed to bridge the modality gap [5, 17, 36, 26].
In the sub-field of time-series anomaly detection, LLMs are increasingly utilised to provide detection capabilities, but also natural language explainability. Architectures like TriP-LLM [47] and MADLLM [37] train specialised adapters to bridge the modality and domain gaps, while AnomalyLLM [21] explores knowledge distillation from LLMs.
Despite these advancements, the direct zero-shot application of LLMs for TSAD remains an emergent field, characterised by a wide array of diverse strategies. While some approaches rely on powerful closed-source models and elaborate prompting [44, 8, 24], others, like LLMAD [23], integrate domain knowledge and databases, converting time-series into tabular or text formats for few-shot explanations. Additional attempts have been made to obtain anomaly positions directly from the LLM via prompting [33]. However, SigLLM [3] demonstrates that using LLMs in a forecasting-based strategy yields far better results than direct questioning.
Alternatively, research in image-based modalities leverages the maturity of Vision-Language Models (VLMs) by transforming time-series into plots or frequency maps [14, 51]. While these visual foundations offer an alternative for cross-modal analysis, this work focuses on the efficiency and granularity of direct time-series-to-token conversions for multivariate time-series, in zero-shot setting.
Methodology
We formalise the TSAD problem in § 3.1 We then detail the tokenisation methods in § 3.2 and the quantification of transferability in § 3.3, as illustrated in Figure 1. Finally, we describe standard adaptations of TSFMs for anomaly detection, used as baselines, in § 3.4.
Figure 1. We adopt a windowed approach, assigning one anomaly score per window. Each window is tokenised step-wise (one token per time-step), processed by a pre-trained LLM, and aggregated into a fixed-size feature vector. An Isolation Forest is then fitted on these features to produce anomaly scores. The resulting performance reflects the suitability of the feature space for anomaly detection and, consequently, the representations transferability to MTSAD. The Raw path corresponds to applying the Isolation Forest directly on the dataset; the Control one on the tokenised dataset.
Problem Formulation
Let be a multivariate time-series of length with variables, with corresponding labels , indicating normal and anomalous measurements. The goal of TSAD is to find a function that assigns an anomaly score to each measurement. Since anomalies are context-dependent, operates on a observation window of the past measurements, i.e., , where . We thus process the time-series as a set of overlapping windows , with . A decision threshold is applied to to obtain predicted labels , which we compare to during testing to assess model performance.
Step-Wise Tokenisation Methods
In an attempt to quantify the zero-shot transferability of sequential knowledge in LLMs to MTSAD, we explore simple, least-disruptive methods of encoding time-series for LLM input. Two main challenges arise: (i) a mismatch between the dimensionality of multivariate time-series and the LLM embedding, and (ii) a more fundamental modality gap between continuous signals and discrete text.
Since subtle local deviations can be important in MTSAD, we prioritise encodings that preserve full temporal detail. We therefore consider step-wise tokenisation functions that map each time-step to a token of dimension (LLM embedding size). We evaluate three alignment strategies.
First, following Zhou et al. [50], we employ Padding, which zero-pads each time-step vector . Second, we utilise Interpolation, which rescales the input by in-filling synthetic dimensions between the original features. These approaches are inherently non-disruptive since .
Finally, to weigh the modality gap against potential information loss, Closest Vocabulary quantises each time-step by mapping interpolated features to their nearest token embeddings via cosine similarity. Unlike learned mappings (Sun et al., [36]), this approach is non-parametric but destructive, as it discretises continuous values into a fixed vocabulary.
Quantifying Anomaly Separability
We use pre-trained LLMs as feature extractors and assess their zero-shot applicability to MTSAD by quantifying the isolation of anomalous representations. The obtained token sequence from a window is passed through the LLM; the processed sequence is finally flattened to obtain the feature vector . After obtaining the set of all window features, we assess the separability of features from normal and anomalous windows.
As no prior knowledge about the nature or structure of anomalies is assumed, we adopt a generic outlier-based perspective: anomalies are treated as observations that deviate from the distribution of normal data. Under this assumption, we use the Isolation Forest (Liu et al., [22]) algorithm to evaluate the suitability of the induced feature space for anomaly detection.
Isolation Forest is an unsupervised method that isolates samples by partitioning the feature space. Let denote an Isolation Forest composed of binary trees, each of depth . Each tree is constructed by recursively partitioning the feature space using random splits along randomly selected dimensions. For the feature vector , whose dimensionality is reduced in practice with Principal Component Analysis, its path length in tree is defined as the number of edges from the root to the terminal node containing . Its expected path length over the forest, , is used to define the anomaly score as:
where is a normalisation constant based on the tree depth.
In this formulation, smaller values of (i.e., shorter paths) correspond to higher anomaly scores, which stems from the core idea of Isolation Forest that anomalous points are rare and structurally distinct, and are therefore isolated more quickly through fewer random splits.
Due to its low parameterisation and computational efficiency, Isolation Forest is particularly well suited for evaluating fixed feature representations without any additional training or optimisation.
Forecasting & Reconstruction Baselines
We identify two main categories of TSFM usage. The first corresponds to task-specific foundation architectures, designed to address time-series challenges such as forecasting, classification, or anomaly detection. These models are typically developed, trained, and evaluated for a specific task and dataset. The second category consists of pre-trained models, trained on large-scale, generic objectives and subsequently transferred to new datasets and tasks.
For TSAD, both approaches generally rely on proxy objectives, most commonly reconstruction or forecasting. In the reconstruction setting, the anomaly score is derived from the Mean Square Error (MSE) between the input window and its reconstruction :
In the forecasting setting, models are given a window and predict the next time-steps for each of the variables. These models typically output a predictive distribution in the form of quantile forecasts (, and ) To exploit this distributional output, the PinBall loss (), a native scoring rule for quantile regression, is usually employed:
Intuitively, observation windows are considered anomalous if they lie in the tails of the predicted distribution, i.e., if they are unlikely under the model’s forecast. The anomaly score is calculated by aggregating the deviations across forecast steps and quantiles, resulting in a scalar value that increases with the level of surprise:
Experiments
Protocol. Evaluating models based solely on the label of the last time-step of each window can be overly restrictive. As the window size increases, anomalies may occur earlier within the window, making this formulation unnecessarily difficult for this study. To relax this constraint, we define the window label as , i.e., a window is labelled anomalous if any of its time steps is anomalous. This choice is a standard procedure in TSAD [49] and only slightly alters the proportion of anomalous samples as a function of . To ensure consistent sample counts across experiments, we adopt a fixed striding strategy (set to the largest window size, 128), which has the additional effect of making each
time-step contribute to at most one window. Adjusted statistics are provided in the supplementary material, but we observed that varying these settings has limited impact on the results (complementary experiments with a unit stride are provided in supplementary).
Evaluation is performed exclusively on the test set, as the training data contains no anomalies, but no label is required by the Isolation Forest. To account for stochasticity in Principal Component Analysis (PCA) and Isolation Forest algorithms, all results are averaged over five random seeds for the pre-trained and random experiments. Results of AnomalyTransformer, TimesNet, and GPT4TS baselines are also averaged across five seeds.
Datasets. We experimented with four common multi-variate time-series anomaly detection datasets. Specifically, they are: Pooled Server Metrics (PSM) [1]; PUMP Fillage (PUMP) [9]; Secure Water Treatment (SWaT) [27]; Water Distribution (WaDi) [2]. Although these datasets primarily cover the industrial domain, they represent core MTSAD benchmarks and encompass diverse types of anomalies. Nevertheless, given the limitations and heterogeneity of MTSAD data [42], we acknowledge that evaluating additional datasets would provide a more comprehensive assessment of the applicability of LLM-derived representations to MTSAD, and defer this to future work. The statistics of the datasets can be found in the supplementary material.
Metrics. To evaluate the anomaly detection performance and thus the models transferability to MTSAD, we consider three main metrics: the Area Under the Receiver Operating Characteristic Curve (AUROC), the Area Under the Area Under the Precision–Recall Curve (AUPR) and the F-score (F) with a threshold set by Youden’s Index. We report F as an indication of in-situ performance, while acknowledging that selecting the threshold based on evaluation across the full test set introduces label dependence. Youden’s Index nevertheless provides a consistent, balanced criterion for inter-model comparison, avoiding F-maximising threshold sweeps or anomaly-ratio-based thresholding, both of which are common in TSAD studies [41].
Model Architectures. We chose to use four open source LLM architectures with available weights in our pipeline: GPT2 [32], Qwen2 (0.5B) [45], Phi-3 [30] and Mistral-7B [18]. In § 4.4, we study two additional Qwen versions: Qwen2 (7B) [45] and Qwen2-VL (7B) [38]. These architectures were selected for their diverse origins, likely distinct pre-training corpora, ease of integration into our current systems, and comparable release timelines.
We complement our study with four TSFM baselines outside of our pipeline: TimesNet [41], GPT4TS [50], Chronos-2 [4] and MOMENT [12]. TimesNet and GPT4TS are reconstruction-based methods, the latter being LLM-based, that are specifically trained on each dataset (Section 4.1). On the other hand, Chronos-2 is already pre-trained on a forecasting task and can be used in a zero-shot setting with a forecasting strategy (Section 4.1), or as a feature extractor (Sec-
tion 4.3). Similarly, MOMENT is pre-trained on a reconstruction objective and can be used in a zero-shot setting with a reconstruction error scoring strategy (Section 4.1). Additionally, we trained AnomalyTransformer (Xu et al., [43]) to provide ground results from a method that is entirely designed and optimised to detect anomalies in time-series.
Model version identifiers are omitted hereafter for conciseness; for all figures, and positive results imply better average performance than the baseline.
Performance of tsfm
We report results for standard TSFM approaches applied to MTSAD in Table 1. TimesNet and GPT4TS are trained in an unsupervised manner using a reconstruction objective, with anomaly scores derived from the reconstruction error (MSE). In contrast, Chronos is used as a pre-trained model, and anomaly scores are computed from the forecasting error using the Pinball loss (). Results are reported for forecasting horizons of 16, 32, 64, and 128.
| Model | PSM | PUMP | SWaT | WaDi | |||||||||
| AUPR | AUROC | F | AUPR | AUROC | F | AUPR | AUROC | F | AUPR | AUROC | F | ||
| Trained MSE | AnomalyTransformer | 68.12±00.40 | 78.97±00.64 | 64.00±00.85 | 22.00±00.07 | 79.56±00.06 | 49.09±00.75 | 69.97±01.27 | 80.88±00.10 | 71.69±02.15 | 29.27±00.01 | 53.03±00.10 | 14.99±00.14 |
| GPT4TS | 52.76±00.11 | 62.05±00.07 | 48.65±00.23 | 07.35±00.00 | 21.02±00.03 | 18.66±00.05 | 10.92±00.02 | 22.12±00.02 | 10.33±00.13 | 41.37±00.00 | 55.82±00.01 | 14.07±00.01 | |
| TimesNet | 51.83±00.10 | 61.55±00.06 | 47.72±00.26 | 07.38±00.01 | 21.15±00.06 | 18.83±00.17 | <u>10.72±00.02</u> | <u>21.83±00.07</u> | <u>09.39±01.46</u> | <u>41.38±00.01</u> | <u>55.95±00.07</u> | <u>14.05±00.01</u> | |
| Pre-Trained L | Chronos () | 44.17 | 57.55 | 44.58 | 09.04 | 27.75 | 12.90 | 09.63 | 26.94 | 02.04 | 06.68 | 48.62 | 13.14 |
| Chronos () | 43.84 | 59.86 | 53.37 | 09.51 | 27.84 | 16.88 | 09.44 | 26.96 | 05.49 | 06.92 | 50.40 | 12.88 | |
| Chronos () | 44.29 | 59.16 | <u>53.10</u> | 10.07 | 31.99 | 20.08 | 10.10 | <u>27.29</u> | 03.26 | 07.60 | 53.61 | 13.55 | |
| Chronos () | 46.72 | 60.65 | 52.13 | 10.26 | 36.74 | <u>21.75</u> | 10.49 | 27.54 | 06.90 | 09.38 | 56.06 | 15.58 | |
| Pre-Trained MSE | MOMENT (Small) | <u>53.36</u> | <u>65.11</u> | 47.71 | 12.50 | <u>43.85</u> | 20.43 | 09.44 | 25.32 | 02.32 | 11.31 | 49.01 | 12.78 |
| MOMENT (Base) | 53.61 | 65.31 | 48.39 | <u>12.44</u> | 43.57 | <u>20.80</u> | 09.39 | 25.25 | 02.39 | 11.11 | 49.21 | 12.39 | |
| MOMENT (Large) | 52.98 | 65.01 | 48.96 | 12.23 | 46.03 | 20.57 | 08.99 | 25.31 | 02.10 | 08.10 | 48.66 | 12.16 |
Table 1. Trained and pre-trained TSFM baselines results with reconstruction (MSE) and forecasting () proxy objectives. AnomalyTransformer is included as a baseline for comparison. Among the remaining results, the best and second-best are shown in bold and <u>underlined</u>, respectively.
TimesNet and GPT4TS yield nearly identical performance, consistent with recent TSAD benchmarks ([31]), and perform similarly to other baselines across most datasets. Chronos and MOMENT largely surpass TimesNet and GPT4TS on PUMP, and achieve the majority of the top AUPR and AUROC results on other datasets. However, AUROC scores stay close to or below 50% across most setting, and noticeably behind AnomalyTransformer on three out of the four studied datasets. These results suggest that standard forecasting and reconstruction objectives in TSFMs, trained or zero-shot, are suboptimal for robust anomaly detection, creating an opportunity to investigate alternative strategies including LLM-driven approaches.
Evaluating MOMENT across its Small, Base, and Large variants (roughly scaled by a factor of three in parameter count) reveals marginal but overall slightly declining performance differences. This suggests that increased model capacity does not necessarily improve performance in TSAD, aligning with our concern that higher reconstruction capability may even degrade detection quality.
Tokenisation Strategies
We explore three tokenisation strategies: Padding, Interpolation, and Closest Vocabulary. For each method, results are averaged across models and datasets, and reported relative to a baseline (Figure 2).
Figure 2. Comparison of 3 step-wise tokenisation methods. We expose the AUROC result differences between the Control path and after the LLMs (dotted bars). For plain bars, results with Padding serve as a baseline and study the differences between those and the ones achieved with Closest Vocabulary and Interpolation. Padding average scores are indicated on the slices sides.
We first compare the performance of an Isolation Forest applied directly to the tokens (Control as a baseline) with the performance obtained after LLM processing (dotted bars). For Padding and Interpolation, LLM processing yields substantial improvements, with gains of up to AUROC points for Mistral and Qwen. In contrast, no improvement is observed for Closest Vocabulary, where performance even slightly degrades. This suggests that the benefit of LLMs emerges most when tokenisation preserves sufficient information, allowing them to meaningfully enrich the input representations.
To compare tokenisation strategies directly, we then take Padding as a baseline and report relative differences (after LLM processing). Closest Vocabulary consistently underperforms, with drops of up to 10 AUROC points for Mistral. Interpolation achieves the best results overall, although the gains remain modest for GPT, Mistral, and Qwen (up to points) and are more pronounced for Phi (up to points).
References
- [1]Ahmed Abdulaal, Zhuanghua Liu, and Tomer Lancewicki. Practical approach to asynchronous multivariate time series anomaly detection and localization. In KDD, pp. 2485–2494. ACM, 2021.DOI
- [2]Chuadhry Mujeeb Ahmed, Venkata Reddy Palleti, and Aditya P. Mathur. WADI: a water distribution testbed for research in the design of secure cyber physical systems. In CySWATER@CPSWeek, pp. 25–28. ACM, 2017.
- [3]Sarah Alnegheimish, Linh Nguyen, Laure Berti-Équille, and Kalyan Veeramachaneni. Can large language models be anomaly detectors for time series? In DSAA, pp. 1–10. IEEE, 2024.DOI
- [4]Abdul Fatir Ansari, Oleksandr Schur, Jaris Kükén, Andreas Auer, Boran Han, Pedro Mercado, Syama Sundar Rangapuram, Huibin Shen, Lorenzo Stella, Xiyuan Zhang, Mononito Goswami, Shubham Kapoor, Danielle C. Maddix, Pablo Guerron, Tony Hu, Junming Yin, Nick Erickson, Prateek Mutalik Desai, Hao Wang, Huzefa Rangwala, George Karypis, Yuyang Wang, and Michael Bohlke-Schneider. Chronos-2: From univariate to universal forecasting. CoRR, abs/2510.15821, 2025.arxiv.org/abs/2510.15821
- [5]Yuxuan Bian, Xuan Ju, Jiangtong Li, Zhijian Xu, Dawei Cheng, and Qiang Xu. Multi-patch prediction: Adapting language models for time series representation learning. In ICML, Proceedings of Machine Learning Research, pp. 3889–3912. PMLR / OpenReview.net, 2024.arxiv.org/abs/2402.04852
- [6]Georgios Chatzigeorgakidis, Konstantinos Lentzos, and Dimitrios Skoutas. Multicast: Zero-shot multivariate time series forecasting using llms. In ICDEW, pp. 119–127. IEEE, 2024.arxiv.org/abs/2405.14748
- [7]Zahra Zamanzadeh Darban, Geoffrey I. Webb, Shirui Pan, Charu Aggarwal, and Mahsa Salehi. Deep learning for time series anomaly detection: A survey. ACM Comput. Surv., 57(1):15:1–15:42, 2025.DOI
- [8]Manqing Dong, Hao Huang, and Longbing Cao. Can llms serve as time series anomaly detectors? CoRR, abs/2408.03475, 2024.arxiv.org/abs/2408.03475
- [9]Cheng Feng and Pengwei Tian. Time series anomaly detection for cyber-physical systems via neural system identification and bayesian filtering. In KDD, pp. 2858–2867. ACM, 2021.DOI
- [10]Shanghua Gao, Teddy Koker, Owen Queen, Tom Hartvigsen, Theodoros Tsiligkaridis, and Marinka Zitnik. Units: A unified multi-task time series model. In NeurIPS, 2024.arxiv.org/abs/2403.00131
- [11]Dong Gong, Lingqiao Liu, Vuong Le, Budhaditya Saha, Moussa Reda Mansour, Svetha Venkatesh, and Anton van den Hengel. Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. In ICCV, pp. 1705–1714. IEEE, 2019.arxiv.org/abs/1904.02639
- [12]Mononito Goswami, Konrad Szafer, Arjun Choudhry, Yifu Cai, Shuo Li, and Artur Dubrawski. MOMENT: A family of open time-series foundation models. In ICML, volume 235 of Proceedings of Machine Learning Research, pp. 16115–16152. PMLR / OpenReview.net, 2024.arxiv.org/abs/2402.03885
- [13]Nate Gruver, Marc Finzi, Shikai Qiu, and Andrew Gordon Wilson. Large language models are zero-shot time series forecasters. In NeurIPS, 2023.arxiv.org/abs/2310.07820
- [14]Zelin He, Sarah Alnegheimish, and Matthew Riemer. Harnessing vision-language models for time series anomaly detection. In AAAI, pp. 21690–21698. AAAI Press, 2026.
- [15]Romain Hermary, Vincent Gaudillière, Abd El Rahman Shabayek, and Djamila Aouada. Removing geometric bias in one-class anomaly detection with adaptive feature perturbation. In WACV, pp. 6612–6622. IEEE, 2025.
- [16]Romain Hermary, Samet Hicsonmez, Dan Pineau, Abd El Rahman Shabayek, and Djamila Aouada. Aster: Latent pseudo-anomaly generation for unsupervised time-series anomaly detection. In International Conference on Pattern Recognition, 2026.
- [17]Yuxiao Hu, Qian Li, Dongxiao Zhang, Jinyue Yan, and Yuntian Chen. Context-alignment: Activating and enhancing llms capabilities in time series. In ICLR. OpenReview.net, 2025.arxiv.org/abs/2501.03747
- [18]Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de Las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Steven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b. CoRR, abs/2310.06825, 2023.arxiv.org/abs/2310.06825
- [19]Yushan Jiang, Zijie Pan, Xikun Zhang, Sahil Garg, Anderson Schneider, Yuriy Nevmyvaka, and Dongjin Song. Empowering time series analysis with large language models: A survey. In IJCAI, pp. 8095–8103. ijcai.org, 2024.arxiv.org/abs/2402.03182
- [20]Yuxuan Liang, Haomin Wen, Yuqi Nie, Yushan Jiang, Ming Jin, Dongjin Song, Shirui Pan, and Qingsong Wen. Foundation models for time series analysis: A tutorial and survey. In KDD, pp. 6555–6565. ACM, 2024.DOI
- [21]Chen Liu, Shibo He, Qihang Zhou, Shizhong Li, and Wenchao Meng. Large language model guided knowledge distillation for time series anomaly detection. In IJCAI, pp. 2162–2170. ijcai.org, 2024.arxiv.org/abs/2401.15123
- [22]Fei Tony Liu, Kai Ming Ting, and Zhi-Hua Zhou. Isolation forest. In ICDM, pp. 413–422. IEEE Computer Society, 2008.
- [23]Jun Liu, Chaoyun Zhang, Jiaxu Qian, Minghua Ma, Si Qin, Chetan Bansal, Qingwei Lin, Saravan Rajmohan, and Dongmei Zhang. Large language models can deliver accurate and interpretable time series anomaly detection. In KDD (2), pp. 4623–4634. ACM, 2025.arxiv.org/abs/2405.15370
- [24]Xingyue Liu, Feizhong Zhou, Hanguang Xiao, Zhipeng Li, Shuai Liu, and Lingling Qian. A survey on large language models for medical time series. Expert Syst. Appl., 311:131364, 2026.DOI
- [25]Kevin Lu, Aditya Grover, Pieter Abbeel, and Igor Mordatch. Frozen pretrained transformers as universal computation engines. In AAAI, pp. 7628–7636. AAAI Press, 2022.arxiv.org/abs/2103.05247
- [26]Luca Masserano, Abdul Fatir Ansari, Boran Han, Xiyuan Zhang, Christos Faloutsos, Michael W. Mahoney, Andrew Gordon Wilson, Youngsuk Park, Syama Sundar Rangapuram, Danielle C. Maddix, and Bernie Wang. Enhancing foundation models for time series forecasting via wavelet-based tokenization. In ICML, Proceedings of Machine Learning Research. PMLR / OpenReview.net, 2025.arxiv.org/abs/2412.05244
- [27]Aditya P. Mathur and Nils Ole Tippenhauer. Swat: a water treatment testbed for research and training on ICS security. In CySWATER@CPSWeek, pp. 31–36. IEEE Computer Society, 2016.DOI
- [28]Nesryne Mejri, Laura Lopez-Fuentes, Kankana Roy, Pavel Chernakov, Enjie Ghorbel, and Djamila Aouada. Unsupervised anomaly detection in time-series: An extensive evaluation and analysis of state-of-the-art methods. Expert Syst. Appl., 256:124922, 2024.DOI
- [29]Marcel Meyer, David Zapata Gonzalez, Sascha Kaltenpoth, and Oliver Müller. Benchmarking time series foundation models for short-term household electricity load forecasting. IEEE Access, 13:218141–218153, 2025.DOI
- [30]Microsoft. Phi-3 technical report: A highly capable language model locally on your phone. CoRR, abs/2404.14219, 2024.
- [31]Xiangfei Qiu, Zhe Li, Wanghui Qiu, Shiyan Hu, Lekui Zhou, Xingjian Wu, Zhengyu Li, Chenjuan Guo, Aoying Zhou, Zhenli Sheng, Jilin Hu, Christian S. Jensen, and Bin Yang. TAB: unified benchmarking of time series anomaly detection methods. Proc. VLDB Endow., 18(9):2775–2789, 2025.
- [32]Alec Radford, Jeffrey Wu, et al. Language models are unsupervised multitask learners. Technical report, 2019.
- [33]Alicia Russell-Gilbert, Alexander Sommers, Andrew Thompson, Logan Cummins, Sudip Mittal, Shahram Rahimi, Maria Seale, Joseph Jabour, Thomas Arnold, and Joshua Church. AAD-LLM: adaptive anomaly detection using large language models. In IEEE Big Data, pp. 4194–4203. IEEE, 2024.
- [34]Bernhard Schölkopf, John C. Platt, John Shawe-Taylor, Alexander J. Smola, and Robert C. Williamson. Estimating the support of a high-dimensional distribution. Neural Comput., 13(7):1443–1471, 2001.pubmed.ncbi.nlm.nih.gov/11440593
- [35]Dimitris Spathis and Fahim Kawsar. The first step is the hardest: pitfalls of representing and tokenizing temporal data for large language models. J. Am. Medical Informatics Assoc., 31(9):2151–2158, 2024.
- [36]Chenxi Sun, Hongyan Li, Yaliang Li, and Shenda Hong. TEST: text prototype aligned embedding to activate llm’s ability for time series. In ICLR. OpenReview.net, 2024.arxiv.org/abs/2308.08241
- [37]Wei Tao, Xiaoyang Qu, Kai Lu, Jiguang Wan, Guokuan Li, and Jianzhong Wang. MADLLM: multivariate anomaly detection via pre-trained llms. In ICME, pp. 1–6. IEEE, 2025.arxiv.org/abs/2504.09504
- [38]Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Kevin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution. CoRR, abs/2409.12191, 2024.arxiv.org/abs/2409.12191
- [39]Frank Wilcoxon. Individual comparisons by ranking methods. Biometrics Bulletin, 1(6):80–83, 1945. doi: 10.2307/3001968.DOI
- [40]Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Rémi Louf, Morgan Funtowicz, and Jamie Brew. Huggingface’s transformers: State-of-the-art natural language processing. CoRR, abs/1910.03771, 2019.arxiv.org/abs/1910.03771
- [41]Haixu Wu, Tengge Hu, Yong Liu, Hang Zhou, Jianmin Wang, and Mingsheng Long. Timesnet: Temporal 2d-variation modeling for general time series analysis. In ICLR. OpenReview.net, 2023.arxiv.org/abs/2210.02186
- [42]Renjie Wu and Eamonn J. Keogh. Current time series anomaly detection benchmarks are flawed and are creating the illusion of progress. IEEE Trans. Knowl. Data Eng., 35(3):2421–2429, 2023.
- [43]Jiehui Xu, Haixu Wu, Jianmin Wang, and Mingsheng Long. Anomaly transformer: Time series anomaly detection with association discrepancy. In ICLR. OpenReview.net, 2022.arxiv.org/abs/2110.02642
- [44]Xiongxiao Xu, Haoran Wang, Yueqing Liang, Philip S. Yu, Yue Zhao, and Kai Shu. Can multimodal llms perform time series anomaly detection? In WWW, pp. 5392–5403. ACM, 2026.arxiv.org/abs/2502.17812
- [45]An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, Guanting Dong, Haoran Wei, Huan Lin, Jialong Tang, Jialin Wang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Ma, Jianxin Yang, Jin Xu, Jingren Zhou, Jinze Bai, Jinzheng He, Junyang Lin, Kai Dang, Keming Lu, Keqin Chen, Kexin Yang, Mei Li, Mingfeng Xue, Na Ni, Pei Zhang, Peng Wang, Ru Peng, Rui Men, Ruize Gao, Runji Lin, Shijie Wang, Shuai Bai, Sinan Tan, Tianhang Zhu, Tianhao Li, Tianyu Liu, Wenbin Ge, Xiaodong Deng, Xiaohuan Zhou, Xingzhan Ren, Xinyu Zhang, Xipin Wei, Xuancheng Ren, Xuejing Liu, Yang Fan, Yang Yao, Yichang Zhang, Yu Wan, Yunfei Chu, Yuqiong Liu, Zeyu Cui, Zhenru Zhang, Zhifang Guo, and Zhihao Fan. Qwen2 technical report. CoRR, abs/2407.10671, 2024.arxiv.org/abs/2407.10671
- [46]Samuel Yoon, Jongwon Kim, Juyoung Ha, and Young Myoung Ko. Patch-based memory gate model in time series foundation model. Transactions on Machine Learning Research, 2026. ISSN 2835-8856.arxiv.org/abs/2509.18751
- [47]Yuan-Cheng Yu, Yen-Chieh Ouyang, and Chun-An Lin. Trip-llm: A tri-branch patch-wise large language model framework for time-series anomaly detection. IEEE Access, 13:168643–168653, 2025.DOI
- [48]Xiyuan Zhang, Ranan Roy Chowdhury, Rajesh K. Gupta, and Jingbo Shang. Large language models for time series: A survey. In IJCAI, pp. 8335–8343. ijcai.org, 2024.arxiv.org/abs/2402.01801
- [49]Qihang Zhou, Shibo He, Haoyu Liu, Jiming Chen, and Wenchao Meng. Label-free multivariate time series anomaly detection. IEEE Trans. Knowl. Data Eng., 36(7):3166–3179, 2024.arxiv.org/abs/2312.11549
- [50]Tian Zhou, Peisong Niu, Xue Wang, Liang Sun, and Rong Jin. One fits all: Power general time series analysis by pretrained LM. In NeurIPS, 2023.arxiv.org/abs/2302.11939
- [51]Jiaxin Zhuang, Leon Yan, Zhenwei Zhang, Ruiqi Wang, Jiawei Zhang, and Yuantao Gu. See it, think it, sort it: Large multimodal models are few-shot time series anomaly analyzers. CoRR, abs/2411.02465, 2024.arxiv.org/abs/2411.02465

