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

Doyun Choi; Dooho Lee; Jaemin Yoo

cs.LGSep 29, 2026

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

Authors: Doyun Choi, Dooho Lee, Jaemin Yoo

Organizations: Seoul National University · KAIST · Nums AI Inc.

Abstract

Unsupervised tabular anomaly detection (TAD) aims to identify anomalous rows in tabular data using normal training samples. While conventional methods rely on dataset-specific training and configuration search, recent tabular foundation models (TFMs) enable zero-shot anomaly detection on unseen datasets via in-context learning. Most TFM-based approaches, however, require anomaly-specific pretraining from scratch, making detection inherently dependent on synthetic TAD-specific priors and costly to update. Some approaches instead repurpose pretrained general-purpose TFMs for TAD to avoid this burden, but rely on computationally expensive formulations with restrictive anomaly inductive biases. In this work, we introduce TaskBridge, a new framework that efficiently repurposes pretrained general-purpose TFMs for unsupervised TAD by constructing virtual supervised tasks that directly recast anomaly detection as supervised in-context inference of TFMs. The resulting virtual tasks induce predictive structures under which normal queries and their target pairs receive high support, whereas anomalies tend to violate the induced structures and receive lower support, providing direct anomaly evidence. Across 790 real-world datasets, TaskBridge consistently outperforms 30 baselines, including state-of-the-art TFM-based approaches, without anomaly-specific TFM pretraining or dataset-specific model optimization.

Figures & tables

Panel of Figure 1 — Figure 1: (Left) Elo scores of TFM-based TAD methods on ODDBench; darker colors indicate better performance. TaskBridge achieves the highest Elo scores for both AUCROC and AUCPR. Full baseline comparisons are provided in Appendix C.1 . (Right) Inference-time comparison between the TabPFN-Extension and TaskBridge , with red indicating that TaskBridge is faster. TaskBridge scales more efficiently as both context length and dimensionality increase.

Figure 2: Overview of TaskBridge . Given a query, it assigns its virtual targets under the designed supervised tasks and distinguishes normal and anomalous queries according to the predictive supports from the pretrained TFM.

Figure 3: Overview of data-adaptive task selection. Candidates are evaluated with a pretrained TFM using nominal held-outs and structural-violation probes, and the most suitable tasks with proper configurations are retained.

Methods	Avg. Rank (AUCROC) $\downarrow$	Avg. Rank (AUCPR) $\downarrow$	ELO (AUCROC) $\uparrow$	ELO (AUCPR) $\uparrow$	Top3 Ratio(%) (AUCROC) $\uparrow$	Top3 Ratio(%) (AUCPR) $\uparrow$
DTE-NP	$9.09_{\pm 0.03}\;(2)$	$9.37_{\pm 0.02}\;(2)$	$1167.7_{\pm 1.0}\;(2)$	$1160.1_{\pm 0.7}\;(2)$	$25.3_{\pm 0.6}\;(2)$	$20.3_{\pm 0.5}\;(4)$
KNN	$9.83_{\pm 0.03}\;(3)$	$10.38_{\pm 0.02}\;(3)$	$1144.7_{\pm 0.9}\;(3)$	$1130.5_{\pm 0.7}\;(3)$	$16.8_{\pm 0.4}\;(8)$	$13.7_{\pm 0.6}\;(10)$
TACTIC-Clean	$12.22_{\pm 0.05}\;(5)$	$11.66_{\pm 0.05}\;(4)$	$1086.1_{\pm 1.4}\;(6)$	$1100.2_{\pm 1.3}\;(4)$	$17.9_{\pm 0.4}\;(7)$	$19.0_{\pm 0.5}\;(6)$
OUTFORMER	$13.32_{\pm 0.05}\;(10)$	$12.19_{\pm 0.02}\;(6)$	$1059.1_{\pm 1.4}\;(10)$	$1087.3_{\pm 0.5}\;(5)$	$19.9_{\pm 0.4}\;(4)$	$21.5_{\pm 0.3}\;(3)$
TaskBridge	$\bm{8.39_{\pm 0.10}\;(1)}$	$\bm{7.82_{\pm 0.06}\;(1)}$	$\bm{1189.5_{\pm 3.2}\;(1)}$	$\bm{1208.1_{\pm 2.2}\;(1)}$	$\bm{44.9_{\pm 1.3}\;(1)}$	$\bm{46.3_{\pm 1.6}\;(1)}$

Table 1: Overall performance on ODDBench, showing the five methods with the highest average AUCROC among 31 baselines including TaskBridge . Results are averaged over five random seeds. Parentheses indicate ranks among all 31 methods, and Total Rank is the mean rank across eight evaluation metrics; six representative metrics are shown. Best, second-best, and third-best results are highlighted in red bold , blue underline , and green , respectively. Our method achieves the strongest overall performance across the reported metrics. Full results are provided in Appendix C.1 .

Figure 4: (Left) Elo scores across three OODBench dataset groups with different context (training data) size (

<1

1

K–

10

K, and

\geq 10

K), showing the five methods with the highest average Elo scores in each group. (Right) Per-dataset AUCROC rank distributions across all baselines; colors follow Figure 1 . TaskBridge achieves the highest Elo score across all dataset groups and the strongest overall rank distribution. Corresponding AUCPR results are provided in Appendix C.1 .

Panel of Figure 5 — Figure 5: Ablation studies of TaskBridge . (Left) Effect of the instantiated normality-anchored virtual tasks. Remove- $i$ denotes removing $i$ tasks from the five task templates introduced in Section 3.3 ; error bars indicate the variation across different removal combinations. The full task family achieves the best average performance. (Right) Effect of data-adaptive task selection. We compare the full procedure with a single held-out split, without intra-task selection, and without inter-task selection. The full selection procedure achieves the best overall performance.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Configuration Axis	Applicable Feature Profiles	Candidates
Number of clusters	All	$\{2,3\}$
Numerical feature transformation	Numerical-only, all mixed variants	$\{\texttt{raw},\texttt{robust}\}$
Number of quantile bins	All	$\{2,3\}$
Number of quantile bins for numerical tokenization	All mixed variants	$\{2,3\}$

Appendix

Table 2: Principal configuration axes used to construct candidate configurations for intra-task selection. Each axis is applied when the corresponding virtual-task instantiation is available for the inferred feature profile. These configurations affect only virtual-task construction; the original observed attributes are preserved for general-purpose TFM’s inference.

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 3: Statistics of the ADBench datasets.

	Methods	Avg. AUCROC $\uparrow$	Avg. Rank (AUCROC) $\downarrow$	Avg. AUCPR $\uparrow$	Avg. Rank (AUCPR) $\downarrow$	ELO (AUCROC) $\uparrow$	ELO (AUCPR) $\uparrow$	Top-3 Ratio (AUCROC) $\uparrow$	Top-3 Ratio (AUCPR) $\uparrow$	Total Rank $\downarrow$
Conventional	DTE-NP	$76.96_{\pm 0.01}\;(2)$	$9.09_{\pm 0.03}\;(2)$	$43.63_{\pm 0.01}\;(2)$	$9.37_{\pm 0.02}\;(2)$	$1167.7_{\pm 1.0}\;(2)$	$1160.1_{\pm 0.7}\;(2)$	$25.3_{\pm 0.6}\;(2)$	$20.3_{\pm 0.5}\;(4)$	$2.25\;(2)$
	KNN	$76.25_{\pm 0.01}\;(3)$	$9.83_{\pm 0.03}\;(3)$	$42.41_{\pm 0.01}\;(5)$	$10.38_{\pm 0.02}\;(3)$	$1144.7_{\pm 0.9}\;(3)$	$1130.5_{\pm 0.7}\;(3)$	$16.8_{\pm 0.4}\;(8)$	$13.7_{\pm 0.6}\;(10)$	$4.75\;(3)$
	FeatureBagging	$75.59_{\pm 0.16}\;(4)$	$12.14_{\pm 0.20}\;(4)$	$39.30_{\pm 0.56}\;(10)$	$13.19_{\pm 0.21}\;(10)$	$1086.9_{\pm 4.8}\;(4)$	$1061.4_{\pm 5.2}\;(10)$	$19.0_{\pm 1.1}\;(6)$	$18.0_{\pm 1.1}\;(8)$	$7.00\;(6)$
	LOF	$74.17_{\pm 0.02}\;(6)$	$12.83_{\pm 0.03}\;(7)$	$39.97_{\pm 0.02}\;(9)$	$12.97_{\pm 0.04}\;(8)$	$1068.9_{\pm 0.9}\;(7)$	$1066.0_{\pm 1.1}\;(8)$	$13.1_{\pm 0.4}\;(10)$	$14.3_{\pm 0.5}\;(9)$	$8.00\;(9)$
	DTE-C	$73.46_{\pm 0.05}\;(9)$	$13.02_{\pm 0.11}\;(8)$	$38.16_{\pm 0.09}\;(11)$	$13.75_{\pm 0.15}\;(12)$	$1063.9_{\pm 3.0}\;(8)$	$1046.7_{\pm 3.6}\;(12)$	$10.9_{\pm 0.6}\;(13)$	$9.4_{\pm 0.4}\;(15)$	$11.00\;(11)$
	DTE-IG	$73.34_{\pm 0.43}\;(11)$	$13.09_{\pm 0.30}\;(9)$	$41.92_{\pm 0.34}\;(6)$	$12.15_{\pm 0.23}\;(5)$	$1061.4_{\pm 7.0}\;(9)$	$1085.1_{\pm 5.6}\;(6)$	$15.5_{\pm 1.3}\;(9)$	$18.5_{\pm 0.8}\;(7)$	$7.75\;(7)$

Appendix

Table 4: Extended results corresponding to Table 1 . Total Rank is computed as the mean rank among 31 methods across eight evaluation metrics: average AUCROC and AUCPR, average AUCROC and AUCPR ranks, Elo scores for AUCROC and AUCPR, and Top-3 Ratios for AUCROC and AUCPR. Our method achieves the strongest overall performance across these metrics.

Figure 6: AUCPR performance comparison on ODDBench. (Extended results corresponding to Figure 4 .) (Left) Elo scores across three dataset groups (

<1

1

K–

10

K, and

\geq 10

K), showing the five methods with the highest average Elo scores in each group. (Right) Per-dataset rank distributions across all baselines; colors follow Figure 1 . TaskBridge also achieves the highest Elo score across all dataset groups with different context size and the strongest overall rank distribution for AUCPR.

Method	AUCROC	AUCPR
TaskBridge	$1188_{\pm 2.6}$	$1207_{\pm 2.8}$
DTE-NP	$1167_{\pm 1.1}$	$1157_{\pm 1.0}$
KNN	$1142_{\pm 0.9}$	$1127_{\pm 0.9}$
TACTIC-Clean	$1103_{\pm 1.2}$	$1106_{\pm 1.4}$
FeatureBagging	$1084_{\pm 5.1}$	$1057_{\pm 5.6}$

Appendix

Table 5: Elo scores on the subset of ODDBench datasets for which all 31 methods successfully produce results. The five methods with the highest AUCROC Elo scores are shown.

Figure 7: Elo score comparison across all baselines on all 790 ODDBench datasets. TaskBridge achieves the highest Elo scores for both AUCROC and AUCPR, demonstrating consistently strong performance across diverse tabular datasets.

Figure 8: Elo score comparison across all baselines on partial ODDBench datasets with context (training data) size below

1

K. TaskBridge achieves the highest Elo scores for both AUCROC and AUCPR, demonstrating strong performance even with limited normality-anchoring evidence.

Figure 9: Elo score comparison across all baselines on partial ODDBench datasets with context (training data) size between

1

K and

10

K. TaskBridge achieves the highest Elo scores for both AUCROC and AUCPR.

Figure 10: Elo score comparison across all baselines on partial ODDBench datasets with context (training data) size of at least

10

K. TaskBridge achieves the highest Elo scores for both AUCROC and AUCPR, maintaining strong performance with large amounts of normality-anchoring evidence.

Remove-1	Single	Localized	Global	Prototype	Distributional
Avg. AUCROC	80.08	79.58	79.98	79.94	80.07
Avg. AUCPR	49.76	49.26	50.19	49.77	49.99

Appendix

Table 6: Ablation results when removing a single virtual task from the set of task templates. Single denotes single-attribute, Localized/Global denote localized/global (subspace) projection, Prototype denotes prototype-based organization, and Distributional denotes distributional extremity.

Remove-2	S+L	S+G	S+P	S+D	L+G	L+P	L+D	G+P	G+D	P+D
Avg. AUCROC	79.35	80.11	79.65	79.72	79.30	78.95	79.13	79.62	79.72	79.60
Avg. AUCPR	47.19	49.16	48.42	48.86	48.57	47.82	48.16	49.17	49.63	48.88

Appendix

Table 7: Ablation results when removing two virtual tasks from the set of task templates. S denotes single-attribute, L/G denote localized/global (subspace) projection, P denotes prototype-based organization, and D denotes distributional extremity.

Remove-3	S+L+G	S+L+P	S+L+D	S+G+P	S+G+D	S+P+D	L+G+P	L+G+D	L+P+D	G+P+D
Avg. AUCROC	78.21	77.61	78.02	79.01	79.58	78.52	78.36	78.64	78.07	79.02
Avg. AUCPR	45.20	42.60	43.85	46.30	47.58	45.62	45.57	47.15	45.18	47.96

Appendix

Table 8: Ablation results when removing three virtual tasks from the set of task templates. S denotes single-attribute, L/G denote localized/global (subspace) projection, P denotes prototype-based organization, and D denotes distributional extremity.

Remove-4	Single	Localized	Global	Prototype	Distributional
Avg. AUCROC	77.12	77.99	76.11	76.15	76.79
Avg. AUCPR	42.33	42.46	36.51	39.01	38.76

Appendix

Table 9: Ablation results when removing four virtual tasks from the set of task templates. For brevity, each variant is labeled by the remaining task. Single denotes single-attribute, Localized/Global denote localized/global (subspace) projection, Prototype denotes prototype-based organization, and Distributional denotes distributional extremity.

	2 Remaining Tasks	3 Remaining Tasks	4 Remaining Tasks	Full TaskBridge
Avg. AUCROC	78.11	79.38	79.89	80.16
Avg. AUCPR	44.95	48.43	49.66	50.43

Appendix

Table 10: Further ablation of inter-task selection by randomly retaining increasing numbers of tasks after intra-task selection for score ensembling. Across all ODDBench datasets, performance improves as more tasks are retained, while the full TaskBridge with inter-task selection achieves the best performance.

Panel of Figure 11 — Figure 11: Inference-efficiency analysis of TaskBridge . (Left, middle) Task-selection time across ODDBench datasets using $3$ (default) and $1$ held-out splits, respectively, as a function of context length and the number of attributes. (Right) Average inference time per test sample after task selection, grouped by context length; TaskBridge maintains millisecond-level inference across all groups. Here, $n$ denotes the number of datasets in each group: $<1{,}000$ , $1{,}000$ – $10{,}000$ , and $>10{,}000$ context length. Overall, TaskBridge preserves efficient inference with TAD-repurposed TFMs.

Figure 12: Robustness to context contamination on ODDBench. AUCROC and AUCPR performance are shown relative to TaskBridge at each contamination level ( TaskBridge

=1

), with its absolute performance reported above each group. p

{k}

denotes the setting in which

k\%

of the context samples are replaced with anomalous samples. TaskBridge remains competitive with TFM-based TAD baselines as contamination increases, while the contamination-specific TACTIC-CONT exhibits the strongest robustness under heavier contamination.

Figure 13: Elo score comparison across all baselines on 226 ODDBench datasets, including four TaskBridge variants with restricted context budgets. Each TaskBridge -

i

K variant limits the context size to

i

K samples. All variants, including the

1

K setting, retain strong performance relative to the other baselines.

	Methods	Avg. AUCROC $\uparrow$	Avg. Rank (AUCROC) $\downarrow$	Avg. AUCPR $\uparrow$	Avg. Rank (AUCPR) $\downarrow$	ELO (AUCROC) $\uparrow$	ELO (AUCPR) $\uparrow$	Top-3 Ratio (AUCROC) $\uparrow$	Top-3 Ratio (AUCPR) $\uparrow$	Total Rank $\downarrow$
Conventional	DTE-NP	$\bm{76.96_{\pm 0.01}\;(1)}$	$\bm{8.90_{\pm 0.02}\;(1)}$	$\bm{43.63_{\pm 0.01}\;(1)}$	$\bm{9.18_{\pm 0.03}\;(1)}$	$\bm{1172.2_{\pm 0.9}\;(1)}$	$\bm{1164.6_{\pm 0.9}\;(1)}$	$\bm{27.9_{\pm 0.5}\;(1)}$	$22.2_{\pm 0.5}\;(4)$	$\bm{1.38\;(1)}$
	KNN	$76.25_{\pm 0.01}\;(2)$	$9.65_{\pm 0.02}\;(2)$	$42.41_{\pm 0.01}\;(4)$	$10.20_{\pm 0.03}\;(2)$	$1149.0_{\pm 0.9}\;(2)$	$1134.8_{\pm 1.0}\;(2)$	$18.6_{\pm 0.6}\;(8)$	$14.9_{\pm 0.5}\;(10)$	$4.00\;(2)$
	FeatureBagging	$75.59_{\pm 0.16}\;(3)$	$11.97_{\pm 0.21}\;(3)$	$39.30_{\pm 0.56}\;(10)$	$13.00_{\pm 0.22}\;(10)$	$1091.1_{\pm 5.0}\;(3)$	$1065.9_{\pm 5.3}\;(10)$	$19.8_{\pm 0.9}\;(6)$	$18.9_{\pm 1.3}\;(8)$	$6.62\;(6)$
	LOF	$74.17_{\pm 0.02}\;(6)$	$12.64_{\pm 0.04}\;(6)$	$39.97_{\pm 0.02}\;(9)$	$12.77_{\pm 0.05}\;(7)$	$1073.6_{\pm 1.0}\;(6)$	$1070.9_{\pm 1.2}\;(8)$	$14.4_{\pm 0.6}\;(10)$	$16.3_{\pm 1.0}\;(9)$	$7.62\;(9)$
	DTE-C	$73.46_{\pm 0.05}\;(9)$	$12.85_{\pm 0.12}\;(7)$	$38.16_{\pm 0.09}\;(11)$	$13.58_{\pm 0.14}\;(12)$	$1068.1_{\pm 3.2}\;(7)$	$1051.0_{\pm 3.5}\;(12)$	$11.3_{\pm 0.7}\;(12)$	$9.7_{\pm 0.6}\;(15)$	$10.62\;(11)$
	DTE-IG	$73.34_{\pm 0.43}\;(11)$	$12.91_{\pm 0.30}\;(8)$	$41.92_{\pm 0.34}\;(6)$	$11.96_{\pm 0.23}\;(4)$	$1065.8_{\pm 7.0}\;(9)$	$1089.8_{\pm 5.5}\;(5)$	$16.3_{\pm 1.5}\;(9)$	$20.0_{\pm 1.1}\;(5)$	$7.12\;(7)$

Appendix

Table 11: Backbone generalization of TaskBridge with TabPFN v2.6. Total Rank is computed as in Table 4 . With this alternative pretrained general-purpose TFM backbone, TaskBridge remains highly competitive among 31 methods.

	Methods	Avg. AUCROC $\uparrow$	Avg. Rank (AUCROC) $\downarrow$	Avg. AUCPR $\uparrow$	Avg. Rank (AUCPR) $\downarrow$	ELO (AUCROC) $\uparrow$	ELO (AUCPR) $\uparrow$	Top-3 Ratio (AUCROC) $\uparrow$	Top-3 Ratio (AUCPR) $\uparrow$	Total Rank $\downarrow$
Conventional	DTE-NP	$\bm{76.96_{\pm 0.01}\;(1)}$	$\bm{8.93_{\pm 0.03}\;(1)}$	$43.63_{\pm 0.01}\;(2)$	$\bm{9.19_{\pm 0.03}\;(1)}$	$\bm{1171.5_{\pm 1.0}\;(1)}$	$\bm{1164.2_{\pm 0.9}\;(1)}$	$\bm{27.4_{\pm 0.6}\;(1)}$	$21.5_{\pm 0.5}\;(4)$	$\bm{1.50\;(1)}$
	KNN	$76.25_{\pm 0.01}\;(2)$	$9.67_{\pm 0.03}\;(2)$	$42.41_{\pm 0.01}\;(5)$	$10.21_{\pm 0.03}\;(2)$	$1148.4_{\pm 1.2}\;(2)$	$1134.5_{\pm 0.8}\;(2)$	$18.7_{\pm 0.6}\;(8)$	$14.8_{\pm 0.7}\;(10)$	$4.12\;(3)$
	FeatureBagging	$75.59_{\pm 0.16}\;(4)$	$11.99_{\pm 0.20}\;(3)$	$39.30_{\pm 0.56}\;(10)$	$13.01_{\pm 0.21}\;(10)$	$1090.7_{\pm 4.7}\;(4)$	$1065.7_{\pm 5.1}\;(10)$	$20.1_{\pm 1.2}\;(6)$	$19.2_{\pm 1.3}\;(8)$	$6.88\;(6)$
	LOF	$74.17_{\pm 0.02}\;(6)$	$12.68_{\pm 0.04}\;(7)$	$39.97_{\pm 0.02}\;(9)$	$12.79_{\pm 0.05}\;(8)$	$1072.6_{\pm 1.1}\;(7)$	$1070.5_{\pm 1.1}\;(8)$	$14.4_{\pm 0.7}\;(10)$	$15.9_{\pm 0.9}\;(9)$	$8.00\;(10)$
	DTE-C	$73.46_{\pm 0.05}\;(9)$	$12.88_{\pm 0.12}\;(8)$	$38.16_{\pm 0.09}\;(11)$	$13.59_{\pm 0.14}\;(12)$	$1067.3_{\pm 3.2}\;(8)$	$1050.7_{\pm 3.6}\;(12)$	$11.3_{\pm 0.7}\;(13)$	$9.8_{\pm 0.4}\;(15)$	$11.00\;(11)$
	DTE-IG	$73.34_{\pm 0.43}\;(11)$	$12.93_{\pm 0.30}\;(9)$	$41.92_{\pm 0.34}\;(6)$	$11.97_{\pm 0.23}\;(5)$	$1065.2_{\pm 7.0}\;(9)$	$1089.4_{\pm 5.5}\;(6)$	$16.5_{\pm 1.1}\;(9)$	$20.2_{\pm 0.9}\;(6)$	$7.62\;(7)$

Appendix

Table 12: Backbone generalization of TaskBridge with TabPFN v3. Total Rank is computed as in Table 4 . TaskBridge remains consistently strong with the alternative pretrained general-purpose TFM backbone, achieving the second-best overall Total Rank among 31 methods.

Explore similar work

Mar 15, 2026

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

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.

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

Ask PDF

Mar 19, 2026

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

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.

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

Ask PDF

May 11, 2026

When Normality Shifts: Risk-Aware Test-Time Adaptation for Unsupervised Tabular Anomaly Detection

Unsupervised tabular anomaly detection methods typically learn feature patterns from normal samples during training and subsequently identify samples that deviate from these patterns as anomalies during testing. However, in practical scenarios, the limited scale and diversity of training data often lead to an incomplete characterization of normal patterns. While test-time adaptation offers a remedy, its isolated focus on test-time optimization ignores the critical synergy with training-phase learning. Furthermore, indiscriminate adaptation to unlabeled test data inevitably triggers anomaly contamination, preventing the model from fully realizing its discriminative capability between normal and anomalous samples. To address these issues, we propose RTTAD, a Risk-aware Test-time adaptation method for unsupervised Tabular Anomaly Detection. RTTAD holistically tackles normality shifts via a synergistic two-stage mechanism. During training, collaborative dual-task learning captures multi-level representations to establish a robust normal prior. During testing, a Test-Time Contrastive Learning (TTCL) module explicitly accounts for adaptation risk by selectively updating the model using high-confidence pseudo-normal samples while constraining anomalous ones. Additionally, TTCL incorporates a k-nearest neighbor-based contrastive objective to refine embedding distributions, thereby further enhancing the model's discriminative capacity. Extensive experiments on 15 tabular datasets demonstrate that RTTAD achieves state-of-the-art overall detection performance.

Wei Huang, Hezhe Qiao, Kailai Zhang +3

Test-Time Adaptation Tabular Anomaly Detection Unsupervised Anomaly Detection Tabular ML Anomaly Detection

Ask PDF