acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Unsupervised tabular anomaly detection (TAD) aims to identify anomalous rows in tabular data using normal training samples and plays an important role in many real-world applications. While conventional approaches rely on dataset-specific model fitting and configuration search, recent prior-data fitted network (PFN)-based tabular foundation models (TFMs) enable zero-shot anomaly detection on unseen datasets via in-context learning (ICL). Most TFM-based approaches, however, require anomaly-specific pretraining from scratch, making detection inherently dependent on synthetic TAD-specific priors and costly to update and extend. 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 with TFMs. The resulting virtual tasks induce predictive structures through the TFM's in-context inference, under which normal queries and their virtual 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 existing TFM-based approaches, without anomaly-specific TFM pretraining or dataset-specific model optimization.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Tabular Anomaly Detection, Prior-data fitted networks, Tabular foundation models, Harnessing foundation model

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_ryuAOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

TaskBridge uses a pretrained tabular foundation model (TFM) for one-class anomaly detection (AD) without updating its weights. From a clean normal reference set, it constructs five families of virtual classification tasks based on attribute prediction, local and global projections, clustering, and distributional extremity. It selects task configurations using held-out normal samples and structural perturbations, then combines calibrated predictive support into an anomaly score. The default backbone is TabICLv2, with framework settings calibrated once on 46 ADBench datasets. Evaluation covers 790 ODDBench datasets, five seeds, and 30 baselines, with additional task ablations, alternative backbones, and context-budget studies. The contribution is a reusable task portfolio that connect supervised in-context prediction to anomaly detection.

Critical strengths
  1. My favorite thing about this paper is the approach. I like that the paper makes a meaningful methodological advance through its portfolio of context derived virtual classification tasks for frozen TFMs. Auxiliary prediction and frozen TFM dependency scoring have precedents in SLAD and TabPFN-OD. The distinctive contribution is the integration of profile aware target construction, predictive support selection, and score normalization. Attribute prediction, projections, prototypes, and extremity probe complementary aspects of normal structure, and task configurations are selected without target anomaly labels or backbone updates. Table 9 supports value of this combination: the single-attribute variant obtains mean area under the ROC curve (AUROC) and precision-recall curve (AUPR) of 77.12% and 42.33%, compared with 80.24% and 50.43% for the full system. This provide evidence of useful innovation beyond one auxiliary prediction task (Sections 3.2-3.5, Appendices A.3-A.6, and Table 9).

  2. I also appreciate the methodologies portability across pretrained backbones and view this as a significant strength. The framework works with TabICLv2, TabPFN v2.6, and TabPFN v3, allowing existing predictive models to be reused for anomaly detection. More importantly, Table 12 compares TaskBridge with the likelihood-based AD extension usin the same TabPFN v3 backbone. TaskBridge improves mean AUROC from 73.61% to 75.91% and mean AUPR from 37.08% to 44.64%. This provides evidence of value beyond the default backbone and supports the method as a practical adaptation framework (Appendix C.6).

Critical weaknesses
  1. The caveat to this portability strength is that TaskBridge's strongest overall performance seems to depend on the backbone. TaskBridge rank first overall with TabICLv2, but fourth with TabPFN v2.6 and second with v3. Both alternatives trail DTE-NP in mean AUROC, although v3 achieves the highest mean AUPR. Portability is well supported, but overall leadership does not persist across backbones. The matched v3 improvement is valuable, whereas Table 11 does not establish lift over a v2.6 extension because the extension baseline uses v3. The paper should distinguish these claims and explain the relevant backbone differences (Tables 4 and 11-12, Appendices B.3 and C.6).

  2. I had a hard time understanding the central virtual task explanation on the first read, and I think a concrete example could really help here. The paper is well organized, but I had to connect Section 3.3 with Appendix A.3 to understand how the task constructions yield anomaly evidence. An anomalous row still receives a virtual label from the same construction rule, so detection depends on the TFM's support for that row-label pair. A worked example would make this distinction easier to follow. The explanation could also separate manually designed templates from dataset adaptive configuration and selection, and distinguish the theory's label stability result from anomaly- elective predictive support (Sections 3.1-3.5 and Appendices A.1 and A.3).

Questions for authors
  1. How does TaskBridge compare with the AD extension and a feature prediction error baseline when the backbone and tuning are matched? A TabPFN v2.6 accuracy comparison and a v3 runtime comparison would help clarify this. Also, along this same line, do you have any intuition for why TabICLv2 perform better? This would be interesting to include.

  2. Does task selection improve over an unfiltered ensemble and random ensembles matched to each dataset's retained task count using the same aggregation? This comparison would distinguish the selection rule's benefit from task complementarity.

  3. As already mentioned, could you walk through one normal row and one anomalous row, showing the task inputs, constructed virtual labels, predictive support, and final anomaly scores? Distinguishing template design from context fitting and task selection, and explaining why low anomalous support is expected would make the central mechanism easier to understand.

8 comments

We thank the reviewer for the constructive feedback. We address the concerns below.

(Q1/W1)

Backbone-Dependent Performance and Possible Explanations.

As reported in the original paper, TaskBridge ranks first with TabICLv2, second with TabPFN v3, and fourth with TabPFN v2.6 in terms of Total Rank. We offer two possible explanations for this backbone-dependent performance:

  1. Backbone-specific configuration search space. Different pretrained TFMs exhibit distinct in-context predictive characteristics. Although we independently calibrated the framework-level hyperparameters for each backbone using ADBench, we employed the same search space initially designed for TabICLv2, including 54 task-selection configurations. This shared search space may not fully accommodate the predictive characteristics of other backbones, such as TabPFN v2.6. Notably, the optimal configurations selected on ADBench differed across backbones, suggesting backbone-specific sensitivity to the framework configurations. Thus, adapting the search space to each backbone's predictive characteristics may further improve performance.
  2. Backbone-specific predictive support across virtual tasks. Different pretrained TFMs may vary in their ability to capture the predictive structures induced by our normality-anchored virtual tasks. Specifically, compared with TabICLv2, TabPFN backbones may assign relatively high predictive support even to anomalous query–target pairs under certain virtual tasks, potentially resulting in weaker discrimination between normal and anomalous samples. This suggests that diversifying the normality-anchored task families to better accommodate different backbone characteristics could be a promising direction.

While these explanations remain hypotheses, they motivate future improvements through backbone-aware configuration calibration and virtual-task diversification. We have also incorporated this discussion into Appendix C.6 of the revised manuscript.

Nevertheless, TaskBridge consistently ranks among the top four of 31 methods across all three backbones. Importantly, it combines competitive AD performance with the ability to repurpose frozen, general-purpose TFMs without anomaly-specific backbone pretraining or dataset-specific parameter updates, offering a practical and transferable approach to TAD.


Additional Experiments.

1. Accuracy Comparison with TabPFN-Extension (TabPFN v2.6)

As requested, we additionally compared TaskBridge against the official TabPFN-Extension using the TabPFN v2.6 backbone. Given the limited computational resources and rebuttal period, we evaluated both methods on 200 randomly sampled ODDBench datasets across five seeds. As shown below, on the 196 datasets with complete results across all five seeds (excluding four with incomplete TabPFN-Extension runs, including out-of-memory errors), TaskBridge substantially outperforms the extension under the matched backbone.

Method (Backbone: TabPFN v2.6)AUROC (%)AUCPR (%)
TaskBridge73.15 ± 0.3240.70 ± 0.21
TabPFN-Extension49.51 ± 0.2217.41 ± 0.08

These results provide additional evidence that TaskBridge's advantage over the likelihood-based extension persists when both methods use the same pretrained TFM backbone.

(Implementation Details) We used tabpfn==8.1.0 and instantiated the TabPFN v2.6 classifier and regressor respectively. They were configured with n_estimators=8, auto_scale_n_estimators=True, softmax_temperature=0.9, fit_mode="fit_preprocessors", and inference_precision="auto", using the native v2.6 preprocessing.

2. Accuracy Comparison with a Feature-Prediction-Error Baseline.

To address the reviewer's request, we implemented FPE-Native, a feature-prediction-error-based anomaly detector using the same TabICLv2 backbone as TaskBridge. FPE-Native predicts each non-constant feature from the remaining features, calibrates prediction errors against held-out normal samples, and aggregates the calibrated errors into anomaly scores. For a fair comparison, we also performed hyperparameter selection for FPE-Native on ADBench.

As shown in Table 3, on the same 200 randomly selected ODDBench datasets across 3 seeds, TaskBridge achieves 79.86% AUCROC and 48.88% AUCPR, outperforming FPE-Native (Full) (78.98% and 45.15%) and (Time-budgeted) (78.96% and 45.25%). Notably, the time-budgeted variant's budget matched TaskBridge’s total task-selection and query-inference time. These results support the effectiveness of our virtual-task construction and data-adaptive task selection over exhaustive feature-wise prediction, even under comparable computational budgets.

Full implementation details and results are provided in Appendix C.8, with a summary in Section 4.3 (lines 471–473) of the revised manuscript.

(Q2)

Task Selection vs. Unfiltered and Matched-Size Random Task Ensembles.

We thank the reviewer for suggesting this comparison. We interpret the request as comparing TaskBridge with (i) an unfiltered ensemble retaining all tasks after intra-task configuration selection and (ii) random task subsets matched to TaskBridge's retained-task count, using the same score aggregation.

As defined in Eq. (12) and detailed in Appendix A.5.2, inter-task selection retains tasks with predictive-support separation ρm,cm⋆≥0.5\rho_{m,c_m^\star}\geq 0.5 under the ADBench-calibrated ALL retention policy (Appendix B.1). We additionally examined the retained-task counts across all 790 ODDBench datasets and three seeds (2,370 dataset–seed evaluations). Notably, all five tasks were retained in 2,368/2,370 cases (99.92%). Filtering occurred only for TimeTrackingIssues (seed 1; 4/5 tasks retained) and AppFraudDetection (seed 2; 3/5 retained).

Thus, for 99.92% of evaluations, TaskBridge produces identical results to both the unfiltered ensemble and random task subsets matched to TaskBridge's retained-task count, as all five tasks are retained. For the two remaining cases, we evaluated all possible task subsets containing the same number of tasks as those retained by TaskBridge, while keeping the intra-task configurations and score aggregation unchanged. The results are shown below.

Dataset (Seed)TaskBridgeRandom-k (Mean)Unfiltered
TimeTrackingIssues (1), k=40.5885 / 0.51640.4331 / 0.47410.3767 / 0.4585
AppFraudDetection (2), k=30.9997 / 0.99930.9996 / 0.99820.9999 / 0.9998

(Values denote AUROC / AUCPR. Random-k averages over all possible task subsets of the same size, including the subset selected by TaskBridge.)

These findings, together with the existing ablations, suggest that TaskBridge's overall ensemble benefit primarily stems from complementary virtual tasks and intra-task configuration selection, rather than frequent inter-task filtering. This is consistent with the task-family and intra-task ablations in Figure 5 of the original paper and the performance improvements with increasing numbers of retained tasks in Table 10. Notably, the existing wo-Inter-Selection ablation in Figure 5 uses a random single-task variant, so its performance gap also reflects the loss of task complementarity. Nevertheless, the TimeTrackingIssues result demonstrates that inter-task filtering can provide substantial benefits when an unsuitable task is identified. We also acknowledge that this benefit is not universal, as illustrated by AppFraudDetection.

(Q3/W2)

Thank you for suggesting a concrete example to clarify the central mechanism. We address your suggestions below.


Template design, context fitting, and selection.

We manually specify five task families targeting complementary aspects of normal structure. Their concrete parameters—including quantile boundaries, category supports, and prototypes—are fitted to the normal context. Using held-out normal samples and structural-violation probes, data-adaptive selection first selects a configuration for each task and then determines which tasks to retain for anomaly scoring. Thus, the templates are predefined, while their instantiation and selection adapt to each dataset.


Concrete walkthrough of TaskBridge.

We provide a real-data walkthrough on WineQuality, one of the ODDBench datasets, tracing a normal and an anomalous query from virtual-label construction through predictive support to final anomaly scores. With seed 0, TaskBridge uses 530 normal context rows and retains five tasks after data-adaptive task selection. For illustration, we select a normal–anomaly pair receiving the identical virtual-label vector [0, 0, 1, 0, 1] across k-means, masked subspace projection, radial binning, global projection, and Reg2Class, respectively. Ground-truth test labels are used only after inference to select this illustrative pair. Figure 3 shows the two queries’ feature values relative to the normal context, their task-wise predictive-support deficits, and their final anomaly scores.

Below, we use the radial task to illustrate label construction and predictive-support scoring, before showing how task-wise scores are aggregated.

(1) Label construction. The selected radial task uses two bins and all 11 numerical features with the raw transform, computing r(x)=∥∣x∥∣2/11r(x)=\||x\||_2/\sqrt{11}. The normal-context median radial statistic, b=39.108734b=39.108734, defines the boundary: label 0 for r(x)<br(x)<b and label 1 otherwise. The normal and anomalous queries have radial statistics 44.1829 and 39.1407, respectively, so both receive label 1.

(2) Predictive support and score calibration. Given the feature vectors and virtually labeled normal context, the TFM assigns predictive support of 0.999999 to label 1 for the normal query and 0.288961 for the anomaly. Calibration compares these values with held-out normal supports. None of the 158 held-out normal supports is as low as the anomaly’s support under the radial task, yielding p=(1+0)/(158+1)≈0.006289p=(1+0)/(158+1)\approx0.006289 and task surprisal −log⁡p≈5.0689-\log p\approx5.0689. The normal query instead receives p=0.867925p=0.867925.

(3) Anomaly scoring. Figure 3 shows how task-wise evidence is aggregated into the final anomaly score. For the anomalous query, the masked-subspace and Reg2Class tasks yield calibrated p-values of 0.118750 and 0.043750, respectively. Their surprisals, 2.1307 and 3.1293, are the anomaly’s two smallest task scores, which low2mean averages. Averaging the two smallest task scores gives (2.1307+3.1293)/2≈2.6300(2.1307+3.1293)/2 \approx \mathbf{2.6300} for the anomaly and (0.1417+0.2076)/2≈0.1746(0.1417+0.2076)/2 \approx \mathbf{0.1746} for the normal query. This example illustrates how identical virtual labels can yield different anomaly scores through context-conditioned predictive support.

We have included the detailed walkthrough in Appendix A.7, with a reference in Section 3 (lines 186–188) of the revised manuscript, to make TaskBridge’s central mechanism easier to follow.


Scope of the theoretical result.

Appendix A.1 establishes that normality anchoring stabilizes virtual-task semantics for normal queries. As stated in Section 3.2, this provides the structural basis for nominal predictive coherence (C1); it does not itself guarantee TFM prediction accuracy or low support for every anomaly. Low anomalous support is expected when a query–target pair violates the predictive relationships inferred from the normal context. Anomalies preserving those relationships may remain supported, motivating complementary task families and data-adaptive selection to improve selective task coverage (C2).

Reviewer_JP7rOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses a frozen general-purpose tabular foundation model (TabICLv2 by default) as a one-class anomaly detector, with no anomaly-specific training. From the normal training rows it builds "virtual" classification tasks whose labels are quantile bins of one feature, bins of a random projection, cluster ids, or bins of a radial or rarity statistic. Task configurations are chosen on held-out normal rows and on corrupted copies of them. A test row scores high when the model gives its virtual label low probability compared with the held-out normals, and the kept tasks are combined with low2mean. On the 790 ODDBench datasets the method has the best average AUROC/AUPRC, rank and Elo of 31 methods; the appendix covers ablations, contamination, context size and two other backbones.

Critical strengths
  1. The evaluation is large, and it can be checked. Per-dataset results for all 31 methods and 5 seeds are released, and recomputing Table 4 from these CSV files gave me the same averages. The gain is unlikely to be noise: TaskBridge beats DTE-NP on 460 datasets and loses on 303 (sign test, p around 1e-8).

  2. Nothing is trained. If the result holds up, the next general-purpose TFM can be swapped in without the new synthetic anomaly priors that FoMo-0D or OUTFORMER would need.

Critical weaknesses
  1. Some of the anomaly evidence bypasses the TFM, and the paper never mentions it. The released config sets outside_context_range_to_invalid to true (Configurations/rules/default.yaml, line 94). In the radial, categorical/mixed projection and rarity tasks, a test row whose statistic exceeds a fence fitted on the context, roughly max(99th percentile, Q3 + 1.5 IQR), gets label -1. So does a row with a category unseen in the context (Utilities/task_helper.py, line 899), even with the flag off. Model/detector.py, line 1543, then scores the row as -log(1e-12), the smallest p-value the task can give, without querying the model. Selection drops these rows (Model/detector.py, line 646), so the rule is invisible during selection but can decide the score at test time. The central claim is that in- context predictive support is the anomaly evidence, so I would need results with this rule disabled. Also undocumented: selection runs the backbone with 1 estimator (rules/default.yaml, line 110) and test scoring with 8 (backends/default.yaml, line 7), which means the calibration set in Eq. 74 and the test scores come from different predictors.
  2. The paper claims more than its results show. The per-dataset numbers do not support "consistently outperforms 30 baselines": the median AUROC gain is 0.31 points over DTE-NP and 0.82 over KNN, and a tail of large wins pulls the mean up. With TabPFN v2.6 as backbone, TaskBridge drops to 4th in Total Rank, behind DTE-NP , KNN and TACTIC-Clean. Novelty is overstated too. GOAD, NeuTraL-AD and ICL (Shenkar & Wolf) already score anomalies with auxiliary prediction tasks built from normal data; some are baselines here, but the related work does not discuss them. What this paper adds is in-context inference and task selection in place of a trained network, and it should say so. Finally, the framework hyperparameters were tuned with labels on 46 ADBench datasets, while the conventional baselines use PyOD defaults.
Questions for authors

Answers to these could change my score:

  1. What are the ODDBench results with the out-of-range rule disabled, covering both the fence and the unseen-category path? How many test anomalies get label -1 in at least one kept task?
  2. What happens when selection and scoring use the same number of estimators?
  3. Could KNN and DTE-NP be given the same ADBench tuning budget?

Clarifications: 4. Table 4 reports 80.24 AUROC for TaskBridge, and Table 10 reports 80.16 for "Full TaskBridge" . Which setting gives which number? 5. Are the 100 private ODDBench datasets among the 790, and how were they obtained? 6. Will you release code for Appendix C.2 to C.6 (TabPFN backends, ablations, contamination, Elo)?

The supplement currently reproduces only the main TabICLv2 run. Suggestions that did not affect my score: wall-clock times for KNN, DTE-NP , FoMo-0D and OUTFORMER, since the runtime comparison covers only TabPFN-Extension; a short look at the 46 datasets where AUROC is below 50; and one worked example of a virtual task in Section 3, because every concrete construction is currently in the appendix.

8 comments

We thank the reviewer for the careful and constructive assessment of our work, including the detailed inspection of our implementation and released results. We also appreciate the thoughtful suggestions and, if accepted, will make every effort to incorporate them into the camera-ready version to further improve the paper. Below, we first address the clarifications, followed by the main questions and corresponding concerns.

(Clarification 4—Results in Tables 4 and 10)

Thank you for pointing out this discrepancy. Table 10 extends the wo-Inter-Selection ablation by randomly retaining different numbers of tasks. The data-adaptive task-selection ablations were conducted with seed 0, whereas Table 4 reports averages over five seeds. This explains the difference between the AUROC values of 80.16 in Table 10 and 80.24 in Table 4.

Our subsequent check of the underlying results also identified an incorrectly reported AUCPR value for Full TaskBridge in Table 10. The correct value is 50.32, replacing 50.43. We have corrected this entry and specified the seed-0 setting in Appendix C.2. (lines 1829-1831) We have also clarified in the experimental settings in Section 4 (lines 411-412) and the Reproducibility Statement (lines 568-569) of the revised manuscript that the five-seed setting applies to the main ODDBench benchmark evaluation.

(Clarification 5—ODDBench Data Availability)

Thank you for asking us to clarify the data provenance. Yes, the reported evaluation includes the 100 datasets intended to remain private. As described in our original reproducibility statement, we obtained all 790 datasets from an earlier version of the repository linked in the OutFormer paper, where they were publicly accessible at the time.

After our paper became public, the ODDBench maintainers informed us that those 100 datasets had been released unintentionally and requested their removal from our anonymous repository. We consulted the Area Chair and removed them with the Area Chair’s approval. We have clarified in the updated Reproducibility Statement (lines 572-577) that the reported results were obtained using all 790 ODDBench datasets, while our anonymous repository provides only the 690 public datasets.

(Clarification 6—Code Availability and Further Improvements)

Thank you for highlighting the need for more complete code coverage. We plan to release the complete experimental code, including the experiments in Appendices C.2–C.6. If accepted, we aim to complete this release during camera-ready preparation to support reproducibility and further research on TFM-based anomaly detection.

(Q1/W1—Out-of-Range Handling)

Thank you for identifying this undocumented implementation detail. We agree that assigning a maximal anomaly score to hard-invalid queries bypasses the TFM and therefore requires validation under our predictive-support formulation.

Following your suggestion, we disabled hard-invalid scoring for both fence violations and unseen-category cases. For categorical Reg2Class only, an unseen target remains marked as invalid to avoid an undefined class lookup; however, the corresponding task is excluded from score aggregation for that query and thus contributes no anomaly evidence. We evaluated this setting with the TabICLv2 backbone on all 790 ODDBench datasets over three seeds.

As shown in Table 6, disabling this rule changes average AUROC and AUCPR by only 0.06 and 0.03 percentage points, respectively. This result shows that the reported performance does not materially depend on hard-invalid scoring and further supports the effectiveness of TaskBridge’s task construction and data-adaptive task selection in deriving anomaly evidence from the TFM’s in-context predictive support.

We also measured the prevalence of this behavior in the original implementation. After averaging the proportion over three seeds for each dataset, 613 of 790 datasets (77.59%) had at most 10% of their test anomalies assigned label=-1 in at least one selected task, and the median dataset-level proportion was only 0.55%. When pooling all test anomalies across datasets, the sample-weighted proportion was 14.54%, averaged over three seeds. Together with the disabled-rule results above, these statistics show that hard-invalid scoring did not materially affect aggregate detection performance, further supporting the robustness of the performance reported in our paper.

We have also documented this implementation detail in Appendix A.3 (lines 1013-1017) of the revised manuscript.

(Q2/W1—Number of Estimators)

Thank you for identifying this undocumented difference and asking us to examine its effect. During intra-task selection, TaskBridge evaluates multiple candidate configurations for each task using common held-out nominal and structural-violation samples. To reduce the cost of these repeated TFM evaluations, we used a single estimator, assuming that it would provide sufficiently reliable relative predictive-support signals for candidate comparison. We have explicitly documented this detail in Appendix B.3 (lines 1583-1590) of the revised manuscript.

Following your suggestion, we compared the original paper setting with two matched settings: one estimator for both selection/calibration and scoring (1/1), and eight estimators for both stages (8/8). We evaluated all three settings using the main TabICLv2 backbone on all 790 ODDBench datasets. Due to computational constraints during the rebuttal period, we report averages over three seeds.

As shown in Table 7, both matched settings yield only marginal differences from the original setting, with absolute changes of approximately 0.41 percentage points or less in mean AUROC and AUCPR. These results indicate limited sensitivity to the estimator settings evaluated here and support our use of a single estimator during task selection to reduce computational cost without materially affecting aggregate detection performance.

(W2—Performance Claims)

Thank you for highlighting the distinction between aggregate performance and dataset-wise consistency. We acknowledge that performance gains vary across datasets and that “consistently outperforms 30 baselines” may be interpreted more broadly than intended. We have revised the wording in the Abstract (lines 27-29), Section 4.1 (lines 417-419), and the caption of Figure 8 (lines 1711-1713) in the updated manuscript to clarify the scope of our claim.

However, our conclusion is not based solely on mean AUROC. With the default TabICLv2 backbone, TaskBridge ranks first among all 31 methods across eight complementary evaluation metrics: average performance, average rank, Elo score, and Top-3 Ratio, each evaluated using AUROC and AUCPR. Average rank captures relative performance within each dataset, Top-3 Ratio measures how frequently a method achieves a leading position, and Elo scores summarize pairwise comparison outcomes across datasets. These complementary results provide evidence beyond mean improvements that may be influenced by a small number of large gains. Together, these findings demonstrate TaskBridge’s effectiveness in tabular anomaly detection and its strong aggregate performance against a broad range of TAD baselines.

In the reported backbone comparison, TaskBridge ranks first with TabICLv2, second with TabPFN v3, and fourth with TabPFN v2.6 in terms of Total Rank. Although it does not rank first with every backbone, it remains among the top four of 31 methods with all three backbones.

Several factors may explain the variation across backbones. The current task-selection configuration grid may not fully accommodate the predictive characteristics of a particular backbone, such as TabPFN v2.6. Backbones may also differ in their ability to infer the predictive structures induced by our virtual task families, and a broader set of task families may better exploit their respective capabilities. Guided by these hypotheses, we plan to investigate the underlying causes of this performance variation in future work.

Nevertheless, TaskBridge combines competitive performance with the ability to repurpose frozen, general-purpose TFMs without anomaly-specific backbone pretraining or dataset-specific parameter updates. This combination can offer practical benefits in real-world applications, enabling strong anomaly detection performance across diverse datasets without the cost of training or fine-tuning a dedicated model for each dataset.

(W2—Novelty)

Thank you for asking us to clarify TaskBridge’s novelty. We understand why TaskBridge may appear similar to GOAD, NeuTraL-AD, and ICL, as these methods also derive anomaly evidence from self-supervised auxiliary tasks. However, the role of these tasks and the mechanism by which they support anomaly detection differ substantially from those in TaskBridge.

GOAD, NeuTraL-AD, and ICL construct transformed views or feature partitions and use self-supervised objectives—such as transformation discrimination or contrastive matching between views or feature subsets—to optimize neural networks on normal data from the target dataset. Anomalies are subsequently scored based on how well samples satisfy these learned relationships in the representation space. Their auxiliary tasks therefore provide training objectives for learning dataset-specific representations of normality.

In contrast, TaskBridge constructs normality-anchored virtual supervised tasks that convert unlabeled normal data into virtually labeled contexts for a frozen, general-purpose TFM. These tasks enable the model to infer predictive structure through supervised in-context learning, without dataset-specific parameter updates. Each query is evaluated through the predictive support assigned to its virtual target under this context-induced structure, and the resulting task-wise evidence is calibrated and aggregated.

Our contribution thus extends beyond replacing a task-trained network with a pretrained model. It addresses how to construct virtual supervised tasks under which normal and anomalous query–target pairs differ in their compatibility with the predictive structure induced by a normal context. TaskBridge develops this formulation through context-driven, normality-anchored task construction, the criteria of nominal predictive coherence and selective task coverage, and complementary task families with data-adaptive selection.

We appreciate the suggestion to clarify this relationship in the main paper. In our revised manuscript, we have added a concise discussion in Section 3.1 (lines 197-203) to explicitly distinguish the role of auxiliary tasks in these prior methods from TaskBridge’s virtual-task formulation and in-context inference mechanism.

Reviewer_7AvHOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TaskBridge, a framework that repurposes pretrained general-purpose tabular foundation models (TFMs) for anomaly detection without anomaly-specific pretraining. It constructs virtual supervised tasks from normal reference data and uses the predictive support assigned by a pretrained TFM to the resulting virtual targets as anomaly evidence.

The framework introduces five complementary virtual-task families and a data-adaptive selection procedure based on held-out normal samples and synthetic structural-violation probes. Experiments on 790 ODDBench datasets show strong performance against 30 baselines, with additional analyses of task selection, inference efficiency, context contamination, and backbone transferability.

Critical strengths

1. Effective reuse of general-purpose TFMs.

The paper presents an interesting approach to anomaly detection by reformulating it as supervised in-context prediction, avoiding anomaly-specific pretraining. This provides a potentially useful and transferable way to leverage advances in general-purpose tabular foundation models.

2. Extensive empirical evaluation.

The evaluation covers 790 datasets and 30 baselines, including recent TFM-based anomaly detectors. The authors also provide ablations of virtual-task families and selection strategies, efficiency comparisons, contamination analysis, and results with alternative TFM backbones. These experiments provide substantial evidence of effectiveness within the evaluated setting.

Critical weaknesses

1. Insufficient justification of anomaly coverage through selected virtual tasks.

TaskBridge assumes that low predictive support on selected virtual supervised tasks provides reliable anomaly evidence. However, low support can also arise from inherently difficult prediction tasks or limitations of the pretrained TFM. Conversely, anomalies that preserve the relationships captured by the selected tasks may remain undetected.

Although the authors introduce five task families and demonstrate their complementary contributions through ablation studies, the paper does not sufficiently analyze which real anomaly mechanisms are captured or missed. In particular, the relationship between computational efficiency and anomaly coverage remains unclear. A comparison with exhaustive feature-wise conditional prediction, preferably under matched computational budgets, would help establish the benefits of the proposed task construction and selection.

2. Unclear generalization of synthetic-probe-based task selection.

TaskBridge selects virtual tasks using held-out normal samples and artificially generated structural-violation probes. Although these probes measure departures from selected properties of the normal distribution, their effectiveness as selection criteria for real anomalies is not fully established.

The authors evaluate the contributions of intra-task and inter-task selection, but do not sufficiently investigate whether the selected tasks remain effective across different anomaly mechanisms or how sensitive selection is to the types of synthetic perturbations. This is important because tasks that distinguish synthetic structural violations may not necessarily identify real anomalies with different characteristics.

3. Dependence on clean normal reference data.

The framework assumes that the reference context consists of normal samples, which limits its applicability to general unlabeled datasets with unknown contamination. The authors acknowledge this limitation and provide contamination experiments, but detection performance degrades under heavier contamination. The method's practical applicability therefore remains most clearly established for clean-reference anomaly detection.

Questions for authors

Q1. Anomaly coverage. How do the selected virtual tasks provide sufficient coverage of diverse anomaly mechanisms? Could the authors analyze detection performance by anomaly type and compare against exhaustive feature-wise prediction under comparable computational budgets?

Q2. Predictive support. How does the method distinguish genuinely anomalous observations from normal samples receiving low predictive support due to unpredictable targets or TFM uncertainty?

Q3. Synthetic probe dependence. How sensitive is task selection to different structural-violation operators? Would the selected tasks remain effective when real anomalies differ substantially from the synthetic perturbations used for selection?

Q4. Practical applicability. Given the dependence on clean normal contexts, what level of contamination can the method tolerate before its advantage over existing anomaly detectors becomes negligible?

6 comments

We thank the reviewer for the careful and constructive assessment of our work. Below, we address each question and the corresponding concerns one by one.

(Q1/W1—Anomaly Coverage)

Thank you for highlighting the need to clarify anomaly coverage. TaskBridge constructs multiple virtual-task families because a single task may not sufficiently distinguish normal from anomalous observations: some anomalies can remain compatible with the particular predictive structure induced by that task.

As described in Section 3.3 and Appendix A.3, our five families encode attribute-wise dependencies, local and global directional structure, prototype organization, and distributional extremity into context-derived virtual targets. The TFM evaluates each query–target pair against the predictive structure induced by the virtually labeled normal context. These constructions provide different perspectives on departures from normality, reducing reliance on any single structural property.

This task diversity is coupled with data-adaptive selection (described in Section 3.4 and Appendix A.5), which aims to retain tasks that provide useful anomaly evidence for each dataset. Task diversification broadens the range of structural departures that the candidate tasks can expose, while adaptive selection identifies suitable tasks for the dataset at hand. The ablations in Section 4.2 and Appendix C.2, including single-task variants and random task selection, support the contribution of this combined design to TaskBridge’s strong anomaly detection performance across diverse real-world datasets.

To further evaluate performance across different anomaly mechanisms, we conducted additional experiments on SynBench, which comprises five anomaly groups constructed from three generative priors with distinct anomaly-generation mechanisms. We randomly sampled 25 datasets per group, yielding 125 datasets, and compared TaskBridge with KNN and DTE-NP, two strong baselines on ODDBench.

As shown in Table 1 and Table 2, TaskBridge achieves the highest mean performance in nine of the ten group–metric combinations, including AUCPR in all five groups. It also ranks first in AUCROC on at least 80% of individual datasets in four groups. Its advantage is smaller for SCM structural anomalies: KNN achieves a slightly higher mean AUCROC, while TaskBridge retains the highest mean AUCPR. These results provide empirical evidence that TaskBridge detects anomalies arising from multiple generation mechanisms, while showing that its relative advantage varies by anomaly group.

We have added the experimental setup, descriptions of the five anomaly groups, and detailed results to Appendix C.7, with a summary in Section 4.3 (lines 484–486) of the revised manuscript.

(Q1/W1—Comparison with Feature-Wise Prediction under Comparable Time Budgets)

To address the requested comparison, we implemented FPE-Native, a feature-prediction-error anomaly detector using TabICLv2 to predict each non-constant feature from all remaining features. Prediction errors are calibrated against held-out normal errors and aggregated into anomaly scores. We evaluate a time-budgeted variant, assigning each dataset–seed run a budget equal to TaskBridge’s task-selection plus query-inference time. After completing calibration and query prediction for each feature, FPE-Native checks the elapsed time and stops once the budget is reached or exceeded, aggregating only fully evaluated features.

As shown in Table 3, on the same 200 randomly selected ODDBench datasets across three seeds, TaskBridge achieves 79.86% AUCROC and 48.88% AUCPR, compared with 78.96% and 45.25% for time-budgeted FPE. Full FPE, which evaluates all non-constant features, achieves 78.98% and 45.15%, respectively. These results support the effectiveness of our virtual-task construction and selection framework relative to both exhaustive feature-wise prediction and its time-budgeted variant.

Full implementation details and results are provided in Appendix C.8, with a summary in Section 4.3 (lines 486–488) of the revised manuscript.

(Q2/W1)

Thank you for raising this concern. We agree that low predictive support can arise from difficult targets or limitations of the pretrained TFM, and that anomalies preserving the relationships captured by the selected tasks may remain undetected. TaskBridge incorporates mechanisms throughout its anomaly detection framework to mitigate these risks.

To directly address these concerns, we introduce two desired properties in Section 3.1: nominal predictive coherence (C1), under which normal query–label pairs receive high support, and selective task coverage (C2), under which support is concentrated on normal rather than deviant inputs. To encourage these properties, we first construct virtual task candidates from structures extracted directly from the normal context (Section 3.2). Under the assumptions in Appendix A.1, this normality anchoring stabilizes the virtual-label semantics for normal queries, providing a structural basis for their compatibility with the context-induced predictive task.

Data-adaptive task selection then evaluates these normality-anchored candidates against the two properties, using the backbone’s predictive support on each dataset. As detailed in Section 3.4 and Appendix A.5, our method assesses whether each candidate assigns consistently high support to held-out normal samples and separates them from structural-violation probes in terms of support. These assessments provide empirical criteria for selecting tasks suitable for anomaly detection. To account for backbone-specific predictive characteristics, we select the framework-level selection and aggregation settings once per backbone using labeled development data (Appendix B.1).

TaskBridge’s anomaly scoring is also designed to mitigate these concerns. It calibrates each query’s support against the held-out normal-support distribution for the same task (Appendix A.6). Thus, each task-wise anomaly score reflects how unusual the query’s support is relative to normal samples. The selected low2mean ensemble conservatively averages the two smallest task-wise scores, reducing the influence of an unusual score from a single task when at least three tasks are retained.

Together, these design choices aim to reduce the risks identified by the reviewer. The task-selection ablations and TaskBridge’s strong overall performance against 30 baselines across 790 ODDBench datasets provide empirical support for the approach, while the failure cases noted above remain possible.

(Q3/W2)

Thank you for raising this concern. The context-calibrated structural-violation probes are not intended to approximate or target specific test-time anomaly mechanisms. As described in Section 3.4 and Appendix A.4, they serve as surrogate structural stress tests for assessing selective task coverage (C2). Specifically, they help identify candidate tasks for which the TFM assigns consistently stronger predictive support to held-out normal samples than to controlled departures from the context-derived normal structure. Thus, the probes guide task selection by testing whether predictive compatibility can provide useful anomaly evidence, rather than by assuming that real anomalies follow the same perturbation mechanisms.

To assess sensitivity to the probe operators, we additionally remove each of the three operators—column shuffling, subset replacement, and scaled jitter—one at a time. As shown in Table 4, across 200 randomly selected ODDBench datasets and three seeds, all variants retain performance close to the full setting: the changes in average AUCROC and AUCPR are within 0.12 and 0.24 percentage points, respectively. The selected task-family sets are also highly stable, with mean Jaccard similarities between 0.9903 and 0.9986 relative to the full setting.

We further evaluate these ablations on 125 SynBench datasets spanning five distinct anomaly-generation mechanisms. As shown in Table 5, performance remains similarly stable across all groups, with changes within 0.34 percentage points in AUCROC and 0.61 percentage points in AUCPR. Together with the ODDBench results on real-world semantic anomalies, these findings indicate that task selection is not strongly dependent on any single probe operator and that the selected tasks remain effective beyond the specific perturbations used during selection.

We have added the experimental details and ablation results to Appendix C.9, with a summary in Section 4.3 (lines 490–492). We also refer readers to this analysis in Section 3.4 (lines 320–321) and Appendix A.4 (lines 1180–1182) of the revised manuscript.

(Q4/W3)

We agree that TaskBridge’s applicability is most clearly established for clean-reference anomaly detection. As explicitly acknowledged in our conclusion, the framework assumes that the reference context consists of normal samples, and contamination can compromise the normality-aware structures used for task construction and selection.

Appendix C.4 evaluates contamination levels of 0%, 3%, 5%, and 7%. Although TaskBridge’s absolute performance decreases, it generally remains competitive with or outperforms the evaluated TFM-based baselines that are not specifically pretrained for contaminated contexts.

To more directly examine when TaskBridge’s relative advantage diminishes, we extend the evaluation to contamination levels of 10%, 15%, 20%, 25%, and 30% on a fixed subset of 94 ODDBench datasets, following the data-selection and contamination protocol in Appendix C.4. We keep the query set fixed across contamination levels and exclude all anomalies reserved for context contamination. All results are averaged over five random seeds.

As shown in Figure 1, TaskBridge’s absolute performance decreases with increasing contamination: its mean AUCROC and AUCPR decline from 79.11% and 74.66% under clean contexts to approximately 63.6% and 55.5% at 10% contamination, and 53.1% and 46.0% at 30%. TACTIC-CONT, which is specifically pretrained for contaminated contexts, outperforms TaskBridge throughout the evaluated 10–30% range. Among baselines without contamination-specific pretraining, TaskBridge remains ahead of FoMo-0D and OutFormer across the evaluated levels, although the gaps become very small at 30%. At contamination levels of 20% and above, its AUCROC is comparable to TabPFN-Extension, while its AUCPR is slightly lower. Its performance is also close to TACTIC-CLEAN, which generally performs better at higher contamination levels.

Thus, under this evaluation protocol, TaskBridge’s relative advantage substantially diminishes in the 20–30% contamination range, with the crossover depending on the baseline and metric. Its absolute performance already deteriorates at lower contamination levels. These findings reinforce the clean-reference scope of the current method. As acknowledged in our conclusion, extending TaskBridge to fully unsupervised settings with unknown context contamination remains an important direction for future work.

Comments on the paper

2 comments

Dear Reviewers,

Please respond to the authors' rebuttal if you haven't done so. Thank you.

Regards,

AC

AuthorsOpenReview ↗

Summary of Our Responses and Manuscript Revisions

Dear Area Chair and Reviewers,

Thank you for your constructive feedback. To facilitate the remaining discussion, we summarize how we have addressed your concerns through additional experiments, clarifications, and manuscript revisions. Detailed answers are available in our individual responses.


Validating anomaly coverage and clarifying the role of context-calibrated probes

(7AvH-Q1/W1) To validate TaskBridge’s anomaly coverage, we clarified how its diverse task families and data-adaptive task selection are designed to broaden coverage across different structural departures from normality. We also added experiments on 125 SynBench datasets spanning five explicitly defined anomaly groups (Appendix C.7). TaskBridge achieves the highest mean performance across almost all group–metric combinations, supporting its effectiveness across diverse anomalies.

(7AvH-Q3/W2) We also clarified the role of context-calibrated structural-violation probes in task selection and added probe-operator ablations on ODDBench and SynBench (Appendix C.9). The results show limited sensitivity to removing individual operators, consistent with the probes’ intended role as surrogate structural stress tests for data-adaptive task selection, without assuming that test-time anomalies follow the same perturbation mechanisms.


Validating TaskBridge’s Detection Performance and Efficiency

We conducted the requested comparisons and analyses to further assess TaskBridge’s leading aggregate performance against 30 baselines on 790 ODDBench datasets.

We first verified that detection performance does not materially depend on hard-invalid scoring or differing estimator counts between selection and scoring (JP7r-Q1,Q2/W1). We also showed that TaskBridge outperforms both exhaustive and time-budgeted feature-prediction-error baselines using the same TFM backbone, supporting the effectiveness of our virtual-task approach to repurposing TFMs (7AvH-Q1/W1, ryuA-Q1/W1; Appendix C.8).

Furthermore, even after tuning the two strongest baselines, KNN and DTE-NP, on the same configuration-tuning datasets (i.e., ADBench) with matched configuration counts, TaskBridge remains first across all aggregate evaluation metrics (JP7r-Q3/W2; Appendix C.10). Comparisons with TabPFN-Extension using the same backbone also demonstrate stronger AD performance and improved inference efficiency (ryuA-Q1/W1).

We also discuss possible reasons for performance variation across backbones, while clarifying that TaskBridge remains competitive with all three evaluated backbones (JP7r-W2, ryuA-Q1/W1; Appendix C.6). Alongside these additional analyses and experiments, we revised the scope of performance claims and clarified experimental settings and implementation details in the revised manuscript (Abstract, Reproducibility Statement, Section 4.1, Appendices A.3 and B.3).


Clarifying TaskBridge’s Anomaly Detection Mechanism and Further Analyses

We clarified how TaskBridge uses context-conditioned predictive support of pretrained TFMs as anomaly evidence, and how task diversity, data-adaptive selection, calibration, and score aggregation help mitigate limitations of individual tasks and pretrained TFMs (7AvH-Q2/W1). We also distinguished virtual supervised tasks, which construct virtually labeled contexts for frozen TFMs, from auxiliary tasks used to train dataset-specific models (JP7r-W2; Section 3.1). A real-data walkthrough now illustrates the complete process of TaskBridge's AD paradigm, from virtual-label construction to final anomaly scores (ryuA-Q3/W2; Appendix A.7).

We further evaluated TaskBridge’s robustness under heavier context contamination, an acknowledged limitation of the current method, identifying where its relative advantage diminishes and clarifying its clean-reference scope (7AvH-Q4/W3). Finally, we revised the inter-task selection ablation discussion based on additional analyses to distinguish the contribution of inter-task filtering from task complementarity (ryuA-Q2 and follow-up; Section 4.2, Figure 5, Appendix C.2).


We also incorporated the relevant clarifications requested by reviewers and refined the wording in the revised manuscript, with references in the main text to the additional appendix analyses. For ease of review, our individual responses include direct links to the relevant figures and tables. The supplementary figure and table PDFs linked in our responses are available together at link.

We hope this summary helps clarify how we have addressed your concerns and where the corresponding revisions can be found. We would appreciate your assessment of the additional evidence and revised manuscript, and welcome any remaining questions before the discussion closes.

Thank you again for your time and thoughtful feedback.

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.