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

Jack Yi Wei; Narges Armanfard

cs.LGMar 19, 2026

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

Authors: Jack Yi Wei, Narges Armanfard

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

Abstract

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.

Figures & tables

Panel of Figure 1 — Figure 1: Each column corresponds to a training regime. The top row shows the training dataset, and the bottom row shows the test set and the anomaly-score landscape from ICLAD after conditioning on training dataset. In the test set, orange diamonds denote ground-truth anomalies and dark circles denote ground-truth normal samples. Darker regions indicate higher anomaly scores.

Figure 2: Overview of the ICLAD framework and synthetic task construction. The top region depicts anomaly generation and the construction of supervision tasks. The bottom region shows the two-stage procedure of ICLAD. Left: prior-fitting on synthetic tasks. Right: inference on real datasets.

Panel of Figure 3 — Figure 3: Feature interaction and t-SNE [ 43 ] plots of normal and anomalous samples. The orange rectangles are anomalies; the blue circles are normal samples.

Panel of Figure 4 — Figure 4: Boxplots of AUC-ROC across 57 datasets. Boxes show the interquartile range (IQR) with medians indicated by the center line and whiskers extending to 1.5 times IQR. Models are ordered by average AUC-ROC and color-coded by family: classical (green), deep learning (blue), and ICLAD (red).

Panel of Figure 5 — Figure 5: Critical difference diagrams of average AUC-ROC ranks. Models are color coded by: classical (green), deep learning (blue) and ICLAD (red)

Figure 6: Semi-supervised performance curves against ratio of labeled anomalies. Left: AUC-ROC Right: AUC-PR

Panel of Figure 7 — Figure 7: Efficiency and ablation analysis. Left Top: average inference runtime (seconds). Left Bottom: average total runtime (training + inference)(seconds). Runtime axes use a log scale. Right: Ablation study of ICLAD variants. Performance are measured in average AUC-ROC across the respective supervision setting.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 8 — Figure 8: Two visual examples of SCM that can be sampled from the SCM prior. Gray circles represent observed variables and white circles are latent variables. $x_{i}$ are selected features and $y$ is the (continuous) target variable.

Hyperparameter	Choices
Share node-wise noise	True, False
Apply blockwise dropout	True, False
Preserve feature order	True, False
Blockwise feature selection	True, False
Activation function	Tanh, ELU, Sigmoid, Identity, Sine, Softplus, Heaviside

Appendix

Table 1: Categorical Hyperparameters and configuration settings for SCMs. Each is sampled independently with uniform probability per dataset.

Hyperparameter	Distribution
MLP weight dropout	0.9 $\text{Beta}(a,b)$ , where $a,b\sim\mathcal{U}(0.1,5.0)$
		Max $\mu$	Min $\tilde{\mu}$	Round	Min
MLP depth	TNLU	8	1	Yes	2
MLP width (hidden nodes)	TNLU	180	5	Yes	4
Number of Causes (Nodes at layer 1)	TNLU	12	1	Yes	1
SCM Node noise Std.	TNLU	0.3	0.0001	No	0.0

Appendix

Table 2: Continuous hyperparameter distributions used in synthetic SCM generation. TNLU(

\mu,\tilde{\mu},\text{round},\text{min}

) denotes log-uniform distribution over parameter

\mu

followed by a truncated log normal distribution with minimal values and potentially rounding the sampled value to the nearest integer.

Hyperparameter	Distribution / Constant
Dataset dimension $d_{in}$	$\mathcal{U}(2,512)$
Support set size $N_{s}$	$\mathcal{U}(5,12000)$
Query set size $N_{q}$	1024
Perturbation Noise Std. ( $\sigma_{j}$ )	$\mathrm{LogUniform}(0.1,10.0)$
Categorical corruption rate $\lambda$	$\mathcal{U}(0,1)$

Appendix

Table 3: Hyperparameter distributions or constants used in task generation and perturbation-based anomalies.

Panel of Figure 9 — Figure 9: Feature interaction and t-SNE plots of normal samples and SCM anomalies.

Panel of Figure 10 — Figure 10: Feature interaction and t-SNE plots of normal and perturbation-based anomaly samples. The orange rectangles are anomalies and the blue circles represent normal samples.

Panel of Figure 11 — Figure 11: Boxplots for one-class setting. Boxes show the interquartile range (IQR) with medians indicated by the center line and whiskers extending to 1.5 times IQR. Models are ordered by average AUC-PR or F1 and color-coded by family: classical (green), deep learning (blue), and ICLAD (red).

Panel of Figure 12 — Figure 12: Boxplots for the unsupervised setting. Boxes show the interquartile range (IQR) with medians indicated by the center line and whiskers extending to 1.5 times IQR. Models are ordered by average AUC-PR or F1 and color-coded by family: classical (green), deep learning (blue), and ICLAD (red).

Panel of Figure 13 — Figure 13: Critical difference diagrams for the one-class setting.

Panel of Figure 14 — Figure 14: Critical difference diagrams for the unsupervised setting.

Figure 15: Semi-supervised performance curves measured by F1 (%) against ratio of labeled anomalies.

Figure 16: Average Total Time (Fit + Predict) in seconds vs. Average AUC-ROC in one-class setting.

Figure 17: Average Total Time (Fit + Predict) in seconds vs. Average AUC-ROC in the unsupervised setting.

Dataset	#Samples	#Feat.	#Anom.	%Anom.	Category
ALOI	49534	27	1508	3.04	Image
annthyroid	7200	6	534	7.42	Healthcare
backdoor	95329	196	2329	2.44	Network
breastw	683	9	239	34.99	Healthcare
campaign	41188	62	4640	11.27	Finance
cardio	1831	21	176	9.61	Healthcare

Appendix

Table 4: Characteristics of classical tabular datasets from ADBench (Part 1).

Dataset	#Samples	#Feat.	#Anom.	%Anom.	Category
WBC	223	9	10	4.48	Healthcare
WDBC	367	30	10	2.72	Healthcare
Wilt	4819	5	257	5.33	Botany
wine	129	13	10	7.75	Chemistry
WPBC	198	33	47	23.74	Healthcare
yeast	1484	8	507	34.16	Biology

Appendix

Table 5: Characteristics of classical tabular datasets from ADBench (Part 2).

Dataset	#Samples	#Feat.	#Anom.	%Anom.	Category
CIFAR10	5263	512	263	5.00	Image
FashionMNIST	6315	512	315	5.00	Image
MNIST-C	10000	512	500	5.00	Image
MVTec-AD	5354	512	1258	23.50	Image
SVHN	5208	512	260	5.00	Image
Agnews	10000	768	500	5.00	NLP

Appendix

Table 6: Characteristics of embedding-based datasets (image and text) from ADBench.

Explore similar work

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
TaskBridge: Bridging Unsupervised Tabular Anomaly Detection and In-Context Learning via Virtual Tasks
Sep 29, 2026Doyun Choi, Dooho Lee, Jaemin YooTabular Anomaly Detection Zero-Shot Anomaly Detection
Seoul National University · KAIST · Nums AI Inc.
LLM as Detector: An In-context Learning Approach for Tabular Anomaly Detection
Aug 19, 2026Tu Anh Hoang Nguyen, Dang Nguyen, Thuc Duy Le +2Tabular Anomaly Detection Anomaly Detection
Applied Artificial Intelligence Initiative (A2I2), Deakin University, Australia · Adelaide University, Australia · Monash University, Australia

Mar 15, 2026J/K move · Enter open · S save

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

Patryk Marszałek, Tomasz Kuśmierczyk, Marek Śmieja

Faculty of Mathematics and Computer Science, Jagiellonian University, Kraków, Poland · Doctoral School of Exact and Natural Sciences, Jagiellonian University, Kraków, Poland

Tabular Anomaly Detection Anomaly Detection Tabular ICL Tabular Foundation Models

Open paper AskPDF arXiv

Anomaly detection for tabular data has been a long-standing unsupervised learning problem that remains a major challenge for current deep learning models. Recently, in-context learning has emerged as a new paradigm that has shifted efforts from task-specific optimization to large-scale pretraining aimed at creating foundation models that generalize across diverse datasets. Although in-context models, such as TabPFN, perform well in supervised problems, their learned classification-based priors may not readily extend to anomaly detection. In this paper, we study in-context models for anomaly detection and show that the unsupervised extensions to TabPFN exhibit unstable behavior, particularly in noisy or contaminated contexts, in addition to the high computational cost. We address these challenges and introduce TACTIC, an in-context anomaly detection approach based on pretraining with anomaly-centric synthetic priors, which provides fast and data-dependent reasoning about anomalies while avoiding dataset-specific tuning. In contrast to typical score-based approaches, which produce uncalibrated anomaly scores that require post-processing (e.g. threshold selection or ranking heuristics), the proposed model is trained as a discriminative predictor, enabling unambiguous anomaly decisions in a single forward pass. Through experiments on real-world datasets, we examine the performance of TACTIC in clean and noisy contexts with varying anomaly rates and different anomaly types, as well as the impact of prior choices on detection quality. Our experiments clearly show that specialized anomaly-centric in-context models such as TACTIC are highly competitive compared to other task-specific methods.

Abstract

Explore similar work

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

TaskBridge: Bridging Unsupervised Tabular Anomaly Detection and In-Context Learning via Virtual Tasks

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