acceptodds
Withdrawn from ICLR 2027

TabMATE: A Multi-Agent Framework for Tabular Anomaly Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In tabular anomaly detection (AD), different anomaly mechanisms favor different detection strategies, making the complementary strengths of heterogeneous detectors valuable for reliable detection. However, this expertise can be lost when detector outputs are uniformly aggregated or reweighted toward consensus, particularly when useful specialists disagree with the majority. We introduce TabMATE, a ular ulti-gent framework that organizes heterogeneous detection specialists into a coordinated eam of xperts through a shared score interface. The framework integrates large language model(LLM) agents as detectors spanning statistical, geometric, isolation-based, and foundation-model approaches, together with a learnable coordination layer that determines their contributions to the final prediction. Alongside the framework, we provide a theoretical analysis of the information limits of label-free coordination and a systematic evaluation on 45 ADBench datasets and four controlled anomaly types. Results show that three confirmed anomalies improve detection over uniform aggregation and the tested consensus-based coordination rules, with learned weights reflecting detector specialization across anomaly types. Further comparisons with alternative coordinators and standalone foundation models clarify the framework’s practical scope, showing that stronger coordination does not necessarily yield a stronger predictor than an individual model under matched label budgets. TabMATE provides a modular foundation for integrating complementary detection expertise and systematically studying how specialist diversity and coordination shape multi-agent tabular AD.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Multi-Agent Systems, Tabular Anomaly Detection, Agent Coordination, Weakly Supervised Learning, Ensemble Learning, Reliable AI

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_okSCOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TabMATE, which treats heterogeneous unsupervised anomaly detectors as "agents" and learns a weight for each detector from a few confirmed anomalies via exponentiated gradient (EG). The final score is a weighted average of the detectors' rank-normalized scores.

Critical strengths
  • The analysis of label-free coordination is convincing: consensus-based rules track agreement among detectors rather than their quality, and thus suppress useful minority specialists.
  • The complementary strengths of detectors across anomaly types are clearly shown.
Critical weaknesses
  • The paper ignores that each unsupervised detector works well only when the data matches its assumption, so the problem is essentially few-label model selection. This explains why (1) weights are per dataset, contradicting the per-sample motivation, and (2) the synthetic setting, one assumption-matched anomaly type per dataset, favors simply selecting the matching detector.
  • The method is essentially a weighted ensemble. By Eq. 3 and Lemma 1, the EG weights reduce to a softmax over each detector's estimated AUC, differing from the "select estimated best" and "AUC reweighting" baselines only in the softmax temperature, which is not ablated.
  • The "multi-agent" and "LLM agents" framing is overstated. The detectors are standard statistical models, they do not interact, and LLMs only appear as optional components in the appendix.
  • A family-balanced uniform baseline is missing. The pool contains several detectors from the same family, so plain uniform averaging, against which all gains are reported, is a weak reference.
  • The method requires confirmed anomalies and falls back to uniform averaging otherwise; how much label information this represents across datasets is not discussed.
Questions for authors

Please consider reporting AUCPR, given the low anomaly ratios in several datasets.

Reviewer_x97AOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes TabMATE, which combines multiple anomaly detectors and uses a few confirmed anomalies to learn their weights. It evaluates the method on 45 datasets and four synthetic anomaly types, comparing different weighting methods and label budgets, and examining robustness and runtime.

Critical strengths

The paper addresses a practical problem: improving anomaly detection with very few labels and the experiments cover 45 datasets and examine different label budgets, feedback errors, and runtime.

Critical weaknesses

The paper does not clearly show what it adds beyond existing feedback-driven ensembles such as AAD and GLAD.And inconsistencies between the text and tables make the main conclusions harder to trust.

Questions for authors

1.The presentation could be more concise. The description of heterogeneous experts, the shared score interface, and modular coordination is repeated across the abstract, introduction, and method section. The calibration-and-freezing procedure is also explained several times. Reducing this repetition would leave more room to clarify the differences from existing ensemble methods. 2.Novelty and the multi-agent formulation. How does TabMATE differ technically from feedback-driven ensembles such as AAD, GLAD, and exponential expert-weighting methods? Beyond terminology, what does the multi-agent formulation add to combining independent detectors with fixed global weights? 3.Consistency of the reported results. Could you resolve the discrepancies between the text and tables? In particular, the claimed improvement from 0.826 to 0.806 appears incorrect, and Table 2 reports a larger gain for mirrored EG than default EG. Please also explain why the three-anomaly EG-10 gains differ across Tables 2, 4, and 14, and whether these corrections affect the conclusions. 4.Evaluation protocol and label budget. How do preprocessing before splitting, calibration on detector-training examples, and hidden-label-based candidate-pool construction affect the results? Please clarify which claims concern having a few confirmed anomalies and which concern making only a few label queries, and what evidence supports performance under a label-blind candidate-pool setup. 5.Weighting rule and theoretical interpretation. Since the final EG weights reduce to exponential weighting by empirical rank separation, what advantage does this particular formulation provide? How sensitive are the results to comparison-set size and weight concentration, and what do the theoretical guarantees establish about the final frozen ensemble’s test performance? 6.Benefit over the strongest standalone model. Tables 14 and 16 suggest that the advantage over TabICL depends on label budget and feedback quality. Is this advantage statistically reliable across datasets, and under what conditions does the ensemble justify its substantially higher total runtime? 7.Reproducibility. Is an anonymous implementation or reproducibility package available? Given the differences between reported results and the missing artifacts noted in the appendix, what materials can be provided to verify the main experiments?

Reviewer_gMbyOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper pools nine classical tabular anomaly detectors and a TabPFN density specialist behind a shared rank-normalized score interface. A few confirmed anomalies, obtained through an oracle-k or query-B protocol, calibrate nonnegative combination weights via an exponentiated-gradient rule; the weights are then frozen and test-time scores are a convex combination of detector scores. The paper also proves that label-free coordination cannot reliably rank detector quality under unrestricted target labelings, diagnoses a failure mode of agreement-based weighting, and evaluates on 45 ADBench datasets and four controlled anomaly types, with matched-budget comparisons against alternative combiners, an expanded pool, and standalone foundation models.

Critical strengths

The most valuable aspect of the paper is its evaluation design rather than its method. The shared score interface with a fixed detector pool cleanly separates detector composition from coordination, so different combination rules can be compared on identical score matrices and identical label budgets. The negative results are also stated with care: the impossibility result for label-free coordination is properly restricted to unrestricted target labelings, and the honest finding that coordination does not beat a single strong foundation model under matched label budgets is more informative than most ensemble papers' claims.

Critical weaknesses
  1. The core problem is that the method is too simple for the claims built around it. Once the terminology is stripped away, the contribution is a constrained form of stacking: rank-normalized detector scores are combined by softmax weights fitted on a handful of labeled examples, which the paper's own Table 2 shows is not significantly better than plain logistic stacking on the same labels (p = 0.16). Learning convex weights over fixed detector outputs from a few labels has been studied for decades in the outlier ensemble literature the paper itself cites, and the exponentiated-gradient update does not change this picture. A coordination rule that is interchangeable with a logistic regression baseline cannot carry the weight of a framework paper.

  2. (!!!) The "multi-agent" framing compounds this by misrepresenting what the paper contains. The title promises a multi-agent framework, but the primary "agents" are ordinary statistical detectors with no agent-like interaction of any kind; the only actual LLM involvement is an appendix study whose results are negative, with LLM coordinators significantly underperforming the numerical baseline and the LLM reporter near chance. The abstract's claim that the framework "integrates large language model agents as detectors" does not match the body of the paper. Relatedly, much of the text, and Figures 1 and 2 in particular, read as machine-generated and insufficiently edited, which undermines confidence in how carefully the paper was prepared.

  3. The empirical record does not rescue the contribution. Several central numbers are internally inconsistent: Section 4.2 reports an increase "from 0.826 to 0.806" with a gain of 0.063 against Table 1's baseline of 0.763; Section 4.3 calls EG the best rule at three anomalies while Table 2 shows mirrored EG at +0.132 versus +0.063; and Section 4.4's conclusion about the individual model's advantage contradicts Table 3a. In addition, classical calibration scores are in-sample and candidate preparation uses hidden labels for stratification and positive-availability fallback, so it is unclear whether the reported gains survive held-out calibration and a fully label-blind candidate pool. Taken together, the paper offers a simple method under an inflated framing, supported by an empirical record whose central claims cannot currently be trusted. I do not see a path by which the authors could address the first two concerns within a revision cycle.

Questions for authors

Can you state precisely what TabMATE provides that logistic stacking over rank-normalized detector scores does not, given the matched-budget results in Table 2? Please also provide corrected versions of the Section 4.2 AUC statement, the oracle-3 comparison in Section 4.3, and the team-versus-individual conclusion in Section 4.4, indicating which numbers are authoritative. Finally, on the same datasets and budgets, what happens when detector calibration scores are held out or cross-fitted and candidate preparation neither checks for hidden positives nor alters prevalence using labels?

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

Related papers

Discussion (0)

Sign in to comment.