ReTabAD: Semantic Benchmark for Tabular AD

• ReTabAD is a benchmark for context-aware tabular anomaly detection that restores semantic metadata such as feature descriptions, measurement units, and domain context.
• It introduces a unified evaluation suite across classical, deep, and LLM-based detectors, showing improved AUROC scores (e.g., from 0.691 to 0.847) by leveraging semantic restoration.
• The framework enhances interpretability by pairing tabular data with JSON metadata to align LLM-selected features with traditional feature-attribution methods.

Searching arXiv for ReTabAD and closely related tabular anomaly-detection benchmarks. ReTabAD is a benchmark for context-aware tabular anomaly detection that restores semantic information typically stripped away in prior tabular AD resources. Introduced as “ReTabAD: A Benchmark for Restoring Semantic Context in Tabular Anomaly Detection” (Yoon et al., 2 Oct 2025), it augments tabular datasets with structured textual metadata such as feature descriptions, measurement units, domain context, and label semantics. The benchmark is motivated by the observation that anomaly detection in tabular domains is often irreducibly context-dependent: anomalousness depends not only on statistical rarity in a feature matrix, but also on the meaning of columns, the interpretation of values, and domain-specific notions of normality and risk. ReTabAD therefore reframes tabular AD from purely distributional modeling toward semantically grounded modeling, formalized as conditioning on metadata through $p(x \mid \mathcal{M})$ rather than only modeling $p(x)$ (Yoon et al., 2 Oct 2025).

1. Conceptual motivation and benchmark rationale

ReTabAD was proposed in response to a limitation in established tabular anomaly-detection benchmarks such as those derived from the DAMI Repository and ADBench: they standardize evaluation over tabularized feature matrices but largely omit semantic context (Yoon et al., 2 Oct 2025). In those prior settings, categorical variables are often integer-coded, textual descriptors are removed or collapsed into embeddings, numerical values may be normalized without preserving their original meaning, and benchmark interfaces typically do not expose dataset documentation. ReTabAD argues that this decontextualization is misaligned with practical anomaly detection, where experts routinely rely on domain semantics to decide whether a record is normal or anomalous (Yoon et al., 2 Oct 2025).

The benchmark’s core premise is that a value may be statistically unusual yet semantically benign, or statistically common yet semantically dangerous. The paper’s illustrative example is a resting heart rate of 200 bpm: as a medically interpreted value it is clearly alarming, whereas in normalized form it may appear merely unusual without conveying clinical significance (Yoon et al., 2 Oct 2025). A similar problem arises for categorical or binary markers whose anomaly relevance depends on domain knowledge rather than surface frequency.

This motivates ReTabAD’s three stated contributions: a benchmark of 20 curated tabular datasets with textual metadata and validated anomaly definitions; a unified evaluation suite spanning classical, deep, and LLM-based anomaly detectors; and a zero-shot LLM baseline that directly uses metadata without task-specific training (Yoon et al., 2 Oct 2025). This suggests the benchmark is intended not only as a dataset resource but also as an intervention in benchmark design: it changes what counts as an input to tabular anomaly detection.

2. Dataset construction, curation, and metadata restoration

ReTabAD contains 20 tabular datasets, each paired with a JSON metadata file (Yoon et al., 2 Oct 2025). The collection is assembled from a curated subset of ADBench datasets together with additional datasets from original sources including the UCI Machine Learning Repository and Kaggle (Yoon et al., 2 Oct 2025). Inclusion required four conditions: a clearly defined domain with ground-truth anomaly labels, manageable size with realistic class imbalance, sufficient documentation to recover reliable semantics, and non-saturated difficulty (Yoon et al., 2 Oct 2025).

The curation process explicitly revisited original data sources to correct erroneous preprocessing from ADBench, verify anomaly definitions against source documentation, recover original feature names and categories, and restore numerical scales where prior normalization was unclear or undocumented (Yoon et al., 2 Oct 2025). This quality-focused strategy contrasts with broader but less semantically faithful benchmark collections.

The 20 datasets span multiple real-world domains. Healthcare contributes seven datasets: cardiotocography, cirrhosis, gallstone, glioma, stroke, vertebral, and wbc. Finance contributes campaign and credit. Other domains include cybersecurity/network backdoor, manufacturing automobile and equip, and scientific or applied domains such as biology, astronomy, geophysics, chemistry, forensics, demographics, environment, and telecommunications (Yoon et al., 2 Oct 2025). The benchmark statistics range from 159 samples in automobile to 50,000 samples in backdoor, census, covertype, and quasar after downsampling; feature counts range from 6 to 42; anomaly ratios range from 0.96% to roughly 33% (Yoon et al., 2 Oct 2025).

The metadata is organized at three levels. Dataset-level descriptions include dataset name, source URL, origin or collection context, and links to repositories or publications. Column-level descriptions specify column name, raw type, logical type, human-readable description, and measurement unit when available. Label-level descriptions define which classes are normal, which are anomalous, and how anomaly labels should be interpreted according to source-grounded semantics (Yoon et al., 2 Oct 2025). ReTabAD therefore pairs raw tabular values with structured metadata, preserving original numerical scales and restoring categorical text values rather than leaving them as arbitrary encodings (Yoon et al., 2 Oct 2025).

3. Formal task setting and evaluation protocol

ReTabAD targets tabular anomaly detection in the one-class or unsupervised setting (Yoon et al., 2 Oct 2025). The benchmark models normality via a probability distribution $\mathcal{P}$ with density $p(x)$, where each instance satisfies $x \in \mathbb{R}^K$ and binary labels satisfy $y \in \{0,1\}$, with 0 denoting normal and 1 anomalous (Yoon et al., 2 Oct 2025). Its conceptual extension is the introduction of semantic metadata $\mathcal{M}$, which includes domain knowledge, feature descriptions, and contextual information about normal behavior. The key modeling objective is stated as

$p(x \mid \mathcal{M}).$

This formulation is central to the benchmark’s identity: anomaly scores are intended to depend on data values and semantic context jointly, not merely on statistical deviation in a context-free feature space (Yoon et al., 2 Oct 2025).

The evaluation split follows a standard one-class protocol. Training uses 50% of normal samples only, while the test set comprises the remaining normal samples together with all anomalous samples (Yoon et al., 2 Oct 2025). Each experiment is repeated five times with different seeds (Yoon et al., 2 Oct 2025). The primary metric is AUROC, with AUPRC reported in appendices and F1@K used for reasoning-alignment analyses (Yoon et al., 2 Oct 2025).

The preprocessing policy is also part of the benchmark definition. Datasets larger than 50,000 are downsampled to 50,000 using seed 42; anomaly proportion is capped at roughly one third; and rows with missing values are removed rather than imputed (Yoon et al., 2 Oct 2025). These choices standardize evaluation while retaining semantic readability and manageable computational cost.

A plausible implication is that ReTabAD is less concerned with maximizing raw dataset count than with controlling semantic validity, anomaly definition fidelity, and benchmark interpretability. That orientation differentiates it from broad tabular AD repositories where semantic reconstruction is secondary.

4. Baseline families and the zero-shot LLM framework

ReTabAD evaluates three classes of baselines: classical anomaly detectors, deep learning methods, and LLM-based methods (Yoon et al., 2 Oct 2025). The classical baselines are KNN, LOF, OCSVM, Isolation Forest, and PCA. The deep baselines are DeepSVDD, REPEN, RDP, RCA, GOAD, NeuTraL, DIF, SLAD, MCM, and DRL. The LLM-related baselines are AnoLLM and the benchmark’s proposed zero-shot framework (Yoon et al., 2 Oct 2025).

AnoLLM is included as a relevant prior LLM-based approach to tabular anomaly detection, but it primarily uses column-name-level context rather than richer semantic metadata (Yoon et al., 2 Oct 2025). ReTabAD’s own LLM framework is explicitly zero-shot and context-aware: it does not rely on task-specific fine-tuning, and instead supplies the model with the row to evaluate, structured metadata, normal statistics from training data, and analysis guidelines (Yoon et al., 2 Oct 2025).

The prompt architecture has three components. The system prompt includes role and task definition, semantic context derived from metadata, and analysis guidelines. The data input formatter serializes each row as text using feature names and values. The output query requests structured JSON of the form

$\{ \text{"anomaly\_score": s, "key\_features": F, "reasoning": e \} ,$

where $s \in [0,1]$ is the anomaly score, $p(x)$0 is a list of key features, and $p(x)$1 is a textual explanation (Yoon et al., 2 Oct 2025). This means the framework evaluates not only predictive ranking but also feature-level rationales and natural-language explanations.

The benchmark defines four prompt variants for ablation. Type A provides only normal statistics with anonymized column names. Type B adds feature descriptions. Type C combines feature descriptions and domain knowledge. Type D, the full-description setting, includes normal statistics, feature descriptions, and domain knowledge (Yoon et al., 2 Oct 2025). Normal statistics are represented using 5th and 95th percentiles from training data; numeric values are rounded to three decimals; raw values are serialized without normalization or discretization by default; batch size for inference is 15; and the system allows up to five retries per query (Yoon et al., 2 Oct 2025). The main experiments use Type D by default (Yoon et al., 2 Oct 2025).

The evaluated LLMs are GPT-4o-mini, GPT-4.1, Claude-3.7-sonnet, Qwen3-235B, and Gemini-2.5-pro, with Gemini-2.5-pro reported as the strongest model in detailed analyses (Yoon et al., 2 Oct 2025).

5. Empirical findings on semantic context, competitiveness, and interpretability

The benchmark’s main empirical result is that semantic context materially improves anomaly detection (Yoon et al., 2 Oct 2025). Across the evaluated LLMs, adding metadata produces an average gain of +7.6 percentage points AUROC when moving from the no-description condition to the full-description condition (Yoon et al., 2 Oct 2025). Gemini-2.5-pro exhibits the largest reported jump, from 0.691 to 0.847 AUROC (Yoon et al., 2 Oct 2025). This is the paper’s central evidence that semantic restoration is not merely cosmetic but directly useful for detection.

In the main benchmark table, the proposed zero-shot metadata-aware LLM baseline achieves average AUROC 0.847 and average rank 4.08, compared with MCM at average AUROC 0.825 and average rank 4.03 (Yoon et al., 2 Oct 2025). The same table reports IForest 0.778, OCSVM 0.803, LOF 0.761, DeepSVDD 0.796, NeuTraL 0.818, SLAD 0.817, DIF 0.742, DRL 0.801, and AnoLLM 0.769 (Yoon et al., 2 Oct 2025). On these results, the zero-shot metadata-aware LLM baseline outperforms the prior LLM baseline and exceeds most classical and deep baselines on average (Yoon et al., 2 Oct 2025).

Performance is especially strong on semantically rich datasets. Reported AUROCs for the zero-shot LLM include 0.895 on census, 0.809 on churn, 0.895 on glioma, 0.755 on vertebral, 0.991 on wine, and 0.961 on quasar (Yoon et al., 2 Oct 2025). However, the gains are not universal: on covertype, backdoor, and gallstone, conventional training-based methods remain stronger or near ceiling (Yoon et al., 2 Oct 2025). This suggests metadata is most useful where semantics are difficult to encode numerically or where categorical and domain-specific meaning is central to anomaly definition.

The benchmark also studies interpretability through feature alignment. For anomalous instance $p(x)$2, the reference top-$p(x)$3 features are defined using supervised XGBoost + SHAP as

$p(x)$4

and alignment with LLM-selected features $p(x)$5 is measured by

$p(x)$6

Representative improvements from Type A to Type D include campaign F1@1 from 0.086 to 0.344, churn F1@1 from 0.133 to 0.288, and glioma F1@1 from 0.009 to 0.551 (Yoon et al., 2 Oct 2025). The paper also reports failure cases where added metadata reduces alignment, including backdoor, covertype, quasar, and stroke (Yoon et al., 2 Oct 2025). Thus the interpretability gains are substantial but not monotone across all domains.

A qualitative example from the cirrhosis dataset further illustrates the benchmark’s interpretive aim: a Type A explanation merely identifies numerical deviation, whereas a Type D explanation states that elevated Prothrombin time indicates compromised liver synthetic function (Yoon et al., 2 Oct 2025). The paper treats this as evidence that semantic context can support expert-style reasoning rather than only improved scoring.

6. Position within tabular AD research, limitations, and common confusions

ReTabAD occupies a specific position in the tabular anomaly-detection literature. It does not primarily introduce a new training-based detector; rather, it introduces a benchmark and evaluation framework for semantically enriched tabular AD (Yoon et al., 2 Oct 2025). In that sense, it plays a role analogous to a benchmark correction or benchmark expansion effort, but focused on semantic restoration rather than temporal realism or industrial feature richness. For comparison, TabReD targets realistic deployment conditions in general tabular ML by emphasizing temporal shift and feature-rich industrial datasets (Rubachev et al., 2024), whereas ReTabAD targets the missing semantic layer in anomaly detection specifically (Yoon et al., 2 Oct 2025). The two are related at the level of benchmark realism, but they address different deficiencies.

The benchmark’s limitations are explicit. It contains 20 datasets, which is smaller than broader resources such as ADBench. Its metadata quality depends on the recoverability and accuracy of original documentation. Strong zero-shot results depend on powerful frontier LLMs, especially Gemini-2.5-pro. Metadata does not improve every dataset equally and may introduce confusion or noise in some cases. The evaluation focuses on one-class anomaly detection with curated anomaly definitions rather than covering all practical AD settings. The ethics discussion also warns that performance on healthcare-related datasets should not be interpreted as clinical readiness (Yoon et al., 2 Oct 2025).

A common confusion concerns naming. ReTabAD should not be conflated with RE-Tab, a distinct framework for TableQA that uses verifiable reward modeling over reasoning trajectories (Kwok et al., 30 Jan 2026). RE-Tab stands for Reward-Enhanced Tabular Agent and addresses interactive table transformation for question answering, not anomaly detection (Kwok et al., 30 Jan 2026). The similarity of names can obscure the fact that the two works belong to different subareas: ReTabAD concerns semantically grounded anomaly detection over tabular datasets, whereas RE-Tab concerns reward-guided reasoning in tool-using TableQA systems (Yoon et al., 2 Oct 2025, Kwok et al., 30 Jan 2026).

More broadly, ReTabAD suggests a methodological shift in tabular AD research. A plausible implication is that future systems may combine statistical anomaly scoring with explicit semantic conditioning, hybrid metadata-aware architectures, or richer reasoning modules that move beyond purely numeric representations. The paper itself identifies future directions including semantic-guided data augmentation, hybrid models combining statistical and semantic signals, richer context-aware tabular AD methods, more interpretable anomaly detection, and metadata-driven preprocessing or foundation-model pipelines (Yoon et al., 2 Oct 2025). Within that trajectory, ReTabAD’s main significance is to define a benchmark substrate on which such methods can be evaluated systematically rather than only anecdotally.