LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection

Tu Anh Hoang Nguyen; Dang Nguyen; Thuc Duy Le; Trung Le; Sunil Gupta

cs.LGAug 19, 2026

LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection

Authors: Tu Anh Hoang Nguyen, Dang Nguyen, Thuc Duy Le, Trung Le, Sunil Gupta

Organizations: Applied Artificial Intelligence Initiative (A2I2), Deakin University, Australia · Adelaide University, Australia · Monash University, Australia

Abstract

Anomaly detection in tabular data is challenging because abnormal samples often arise as violations of cross-feature dependencies rather than simple marginal deviations. Existing detectors rely on geometric or reconstruction signals, while prior LLM-based approaches mainly fine-tune LLMs with normal samples or generate synthetic anomalies. We propose LLM-Detector, a framework that utilizes the in-context learning capacity of LLMs for structured, prompt-conditioned scoring synthesis, enabling LLMs to derive anomaly detection logic from structured normal-state knowledge. Specifically, normal training data are converted into statistical summaries, causal dependencies, and distilled prototypes that are organized into a prompt for code generation. The resulting scoring engine evaluates statistical deviation, structural inconsistency, and density-based abnormality then computes an anomaly score for each test sample. We evaluate LLM-Detector on 24 tabular datasets, comparing against 15 SOTA baselines. Results show consistent improvements across both mixed-type and continuous-only settings. Moreover, this design eliminates the need for LLM fine-tuning or neural network training, reducing computational cost and enabling practical anomaly detection in real-world tabular systems.

Figures & tables

Figure 1: Comparison of LLM-based TAD methods across categorical handling, computational efficiency, and causal modeling.

Figure 2: Our framework LLM-Detector . It has two phases: (1) Knowledge-to-Code extracts statistical profiles, causal structure, and distilled prototypes from normal data to construct a structured prompt, and (2) Code-to-Score generates a scoring engine executed on test samples to compute anomaly scores.

Figure 3: Prototype distillation via K-Means and nearest-neighbor (NN) mapping.

Dataset	$N$	$d$	$d_{cat}$	$d_{con}$	$N_{a}$
ACD	10,000	4	3	1	490 (5%)
Bank	41,188	10	10	0	4,640 (11%)
CMC	1,473	8	8	0	29 (2%)
DAMRE	1,000	5	2	3	100 (10%)
Fraud	151,112	8	6	2	14,151 (9%)
Lymp	148	18	15	3	6 (4%)

Table 1: Statistics for 12 mixed-type datasets.

d_{cat}

d_{con}

, and

N_{a}

denote the numbers of categorical features, continuous features, and anomalies. Lymp (Lymphography), ACD (Cybersecurity), and Fraud (Fraud ecommerce) are from [ 22 , 5 , 7 ] while SPD (Spyware-Attacks), DAMRE (Damage-Report), OS (OS-Kernel), SMD (Smart-Meter), and Vifd (Vehicle insurance) are from Kaggle.

	ACD	Bank	CMC	DAMRE	Fraud	Lymp	NHIS	OS	Seismic	SMD	SPD	Vifd	AVG
Iforest	0.5030	0.5118	0.5061	0.6901	0.4800	0.7050	0.6573	0.7716	0.7058	0.8127	0.4837	0.5067	0.6112
KNN	0.4943	0.5124	0.5779	0.9250	0.5005	0.8705	0.6748	0.9330	0.7445	0.9237	0.4919	0.5227	0.6809
PCA	0.5059	0.5152	0.5101	0.8857	0.4538	0.8560	0.6498	0.8475	0.7041	0.9340	0.4786	0.4931	0.6528
ECOD	0.5086	0.5134	0.5394	0.8442	0.4631	0.8545	0.6537	0.7686	0.6980	0.9114	0.4915	0.5028	0.6458
DeepSVDD	0.4774	0.4996	0.5715	0.9162	0.5015	0.8685	0.6626	0.7957	0.6939	0.6864	0.4848	0.4993	0.6381
GOAD	0.4969	0.5139	0.4833	0.8726	0.5001	0.8764	0.7151	0.9117	0.7193	0.1545	0.4910	0.4991	0.6028

Table 2: AUC-ROC scores for all methods on 12 mixed-type datasets. Bold and underline indicate the best and second-best methods. Standard deviations are reported in Appendix 2 .

Figure 4: AUC-ROC comparison on 12 continuous datasets across all 15 methods. The color scheme is: yellow (proximity-based), green (distribution-based), blue (boundary-based), purple (reconstruction/generation-based), navy (LLM-based), red (ours).

Figure 5: Ablation study of three types of knowledge in the description component on 12 mixed-type datasets.

Figure 6: Sensitivity to three LLM backbones (DeepSeek, GPT, and Gemini) over 12 mixed-type datasets.

Figure 7: Sensitivity to the number of distilled samples over 12 mixed-type datasets.

Figure 8: Sensitivity to causal discovery methods over 12 mixed-type datasets.

Figure 9: Computational efficiency comparison across all 24 datasets.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	$N$	$d$	$d_{con}$	$N_{a}$
Annthyroid	7,200	6	6	534 (7%)
BreastW	699	9	9	241 (34%)
Ecoli	336	7	7	9 (3%)
Shuttle	49,097	9	9	3,511 (7%)
Thyroid	3,772	6	6	93 (2%)
Wine	130	13	13	71 (55%)

Appendix

Table 3: Statistics for 12 continuous-only datasets.

Dataset	Iforest	PCA	KNN	ECOD	DeepSVDD	RCA	SLAD	GOAD	NeuTral	ICL	DTE	REPEN	AnoGAN	AnoLLM	LLM-DAS	LLM-Detector
Annthyroid	0.9074 (0.011)	0.8298 (0.024)	0.8088 (0.003)	0.7883 (0.002)	0.7335 (0.014)	0.7160 (0.001)	0.7538 (0.008)	0.5521 (0.002)	0.7766 (0.022)	0.8396 (0.019)	0.9763 (0.002)	0.7316 (0.016)	0.6361 (0.155)	0.9153 (0.001)	0.8330 (0.016)	0.9117 (0.013)
BreastW	0.9926 (0.002)	0.9912 (0.003)	0.9930 (0.003)	0.9894 (0.002)	0.9719 (0.004)	0.9889 (0.003)	0.9896 (0.004)	0.9932 (0.002)	0.9848 (0.003)	0.9915 (0.003)	0.9879 (0.005)	0.9520 (0.021)	0.9283 (0.090)	0.9911 (0.002)	0.9845 (0.002)	0.9673 (0.018)
Ecoli	0.8126 (0.005)	0.8358 (0.008)	0.8785 (0.002)	0.7514 (0.006)	0.8525 (0.016)	0.8844 (0.028)	0.7755 (0.010)	0.8153 (0.006)	0.8645 (0.007)	0.8925 (0.020)	0.8182 (0.005)	0.8424 (0.044)	0.7650 (0.023)	0.8570 (0.013)	0.8241 (0.015)	0.8787 (0.012)
Shuttle	0.9962 (0.000)	0.9936 (0.001)	0.9989 (0.000)	0.9930 (0.000)	0.9949 (0.000)	0.9954 (0.000)	0.9981 (0.000)	0.9883 (0.001)	0.9977 (0.002)	0.9995 (0.000)	0.9976 (0.000)	0.9938 (0.002)	0.9906 (0.002)	0.9984 (0.000)	0.9952 (0.001)	0.9915 (0.001)
Thyroid	0.9903 (0.001)	0.9839 (0.000)	0.9761 (0.001)	0.9787 (0.001)	0.8501 (0.006)	0.9712 (0.001)	0.9485 (0.011)	0.6844 (0.024)	0.8459 (0.017)	0.9827 (0.009)	0.9909 (0.001)	0.9295 (0.017)	0.9606 (0.021)	0.9890 (0.001)	0.9776 (0.002)	0.9617 (0.017)
Wine	0.9355 (0.013)	0.9972 (0.000)	0.9931 (0.002)	0.7437 (0.009)	0.9665 (0.023)	0.9551 (0.007)	0.9917 (0.011)	0.9955 (0.000)	0.9937 (0.001)	0.9660 (0.036)	0.9922 (0.004)	0.3094 (0.067)	0.9890 (0.005)	0.9955 (0.003)	0.9925 (0.001)	0.9853 (0.020)

Appendix

Table 4: AUC-ROC (standard deviation) on each dataset ( higher is better ).

Dataset	Iforest	PCA	KNN	ECOD	DeepSVDD	RCA	SLAD	GOAD	NeuTral	ICL	DTE	REPEN	AnoGAN	AnoLLM	LLM-DAS	LLM-Detector
Annthyroid	0.5512 (0.016)	0.4794 (0.025)	0.4394 (0.005)	0.3882 (0.001)	0.4288 (0.025)	0.3733 (0.006)	0.4051 (0.003)	0.2303 (0.003)	0.4244 (0.013)	0.5568 (0.023)	0.7840 (0.011)	0.3377 (0.014)	0.2903 (0.109)	0.6036 (0.002)	0.4644 (0.017)	0.4819 (0.096)
BreastW	0.9627 (0.006)	0.9599 (0.010)	0.9655 (0.008)	0.9461 (0.012)	0.9336 (0.009)	0.9599 (0.007)	0.9571 (0.010)	0.9641 (0.007)	0.9613 (0.009)	0.9558 (0.002)	0.9585 (0.009)	0.9170 (0.036)	0.8880 (0.106)	0.9613 (0.004)	0.9557 (0.010)	0.9392 (0.019)
Ecoli	0.7037 (0.052)	0.7778 (0.000)	0.7408 (0.052)	0.3703 (0.052)	0.7778 (0.000)	0.5926 (0.052)	0.4444 (0.091)	0.5556 (0.091)	0.6667 (0.091)	0.6297 (0.105)	0.7408 (0.052)	0.5926 (0.262)	0.2592 (0.139)	0.7778 (0.000)	0.4815 (0.210)	0.6667 (0.157)
Shuttle	0.9707 (0.003)	0.9572 (0.000)	0.9766 (0.001)	0.9169 (0.001)	0.9786 (0.003)	0.9694 (0.001)	0.9747 (0.000)	0.9625 (0.001)	0.9840 (0.001)	0.9801 (0.002)	0.9736 (0.001)	0.9677 (0.001)	0.9546 (0.005)	0.9809 (0.000)	0.9777 (0.000)	0.9569 (0.003)
Thyroid	0.8100 (0.018)	0.7383 (0.013)	0.6380 (0.005)	0.6344 (0.000)	0.4982 (0.071)	0.6129 (0.009)	0.6416 (0.052)	0.4158 (0.018)	0.2796 (0.046)	0.7240 (0.020)	0.8100 (0.010)	0.3764 (0.040)	0.5950 (0.106)	0.7957 (0.000)	0.6237 (0.018)	0.6308 (0.082)
Wine	0.9155 (0.020)	0.9812 (0.007)	0.9765 (0.007)	0.8169 (0.020)	0.9342 (0.033)	0.9530 (0.007)	0.9812 (0.018)	0.9718 (0.000)	0.9718 (0.000)	0.9436 (0.042)	0.9671 (0.007)	0.6291 (0.027)	0.9671 (0.018)	0.9812 (0.007)	0.9577 (0.012)	0.9765 (0.013)

Appendix

Table 5: F1-score (standard deviation) on each dataset ( higher is better ).

Figure 10: Full code-generation prompt template used in the Knowledge-to-Code stage. The prompt includes role definition, dataset context, task description, analytical instructions, and expected output format, while replacing full-dataset injection with three compact normal-state knowledge components: feature distributions, causal knowledge, and distilled samples. The distilled sample block is truncated for space efficiency; the actual setup uses 100 distilled samples.

Explore similar work

ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes
Mar 19, 2026Jack Yi Wei, Narges ArmanfardTabular Anomaly Detection Tabular Foundation Models
Department of Electrical and Computer Engineering, McGill University, Canada · Mila - Quebec Artificial Intelligence Institute, Montreal, Canada
TACTIC for Navigating the Unknown: Tabular Anomaly deteCTion via In-Context inference
Mar 15, 2026Patryk Marszałek, Tomasz Kuśmierczyk, Marek ŚmiejaTabular Anomaly Detection Anomaly Detection
Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland · Doctoral School of Exact and Natural Sciences, Jagiellonian University, Kraków, Poland
Enhancing Tabular Anomaly Detection via Pseudo-Label-Guided Generation
Apr 20, 2026Wei Huang, Yuxuan Xiong, Hezhe Qiao +3Tabular Anomaly Detection Pseudo-Labeling

Mar 19, 2026J/K move · Enter open · S save

ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes

Jack Yi Wei, Narges Armanfard

Department of Electrical and Computer Engineering, McGill University, Canada · Mila - Quebec Artificial Intelligence Institute, Montreal, Canada

Tabular Anomaly Detection Tabular Foundation Models Unsupervised Anomaly Detection Tabular ICL

Open paper AskPDF arXiv

Anomaly detection on tabular data is commonly studied under three supervision regimes, including one-class settings that assume access to anomaly-free training samples, fully unsupervised settings with unlabeled and potentially contaminated training data, and semi-supervised settings with limited anomaly labels. Existing deep learning approaches typically train dataset-specific models under the assumption of a single supervision regime, which limits their ability to leverage shared structures across anomaly detection tasks and to adapt to different supervision levels. We propose ICLAD, an in-context learning foundation model for tabular anomaly detection that generalizes across both datasets and supervision regimes. ICLAD is trained via meta-learning on synthetic tabular anomaly detection tasks, and at inference time, the model assigns anomaly scores by conditioning on the training set without updating model weights. Comprehensive experiments on 57 tabular datasets from ADBench show that our method achieves state-of-the-art performance across three supervision regimes, establishing a unified framework for tabular anomaly detection.

Abstract

Explore similar work

ICLAD: In-Context Learning for Unified Tabular Anomaly Detection Across Supervision Regimes

TACTIC for Navigating the Unknown: Tabular Anomaly deteCTion via In-Context inference

Enhancing Tabular Anomaly Detection via Pseudo-Label-Guided Generation