SPOT-FM: A Zero-Shot Hypothesis Test for Tabular Anomalies via In-Context Synthetic Contrasts

We formulate zero-shot tabular anomaly detection as a composite hypothesis test: H₀ states that a query row is drawn from a reference distribution; the alternative is defined by a learned sampler law. We present SPOT-FM, a frozen Tab2D transformer whose classification logit serves as a learned, data-adaptive test statistic, instantiated from a user-supplied reference set with no per-dataset training. The test's alternative hypothesis is constructed via minimal-subset perturbation: only a μ-fraction of features are altered per synthetic contrast, preserving inter-feature structure while producing controlled departures from the empirical quantile box. We prove a data-dependent power lower bound against this alternative (Thm. 1), establish exchangeability of the test statistic under reference-set permutations (Thm. 2), and decompose deployment risk into pre-training fit, task shift, and a sampler-label residual that names where zero-shot transfer fails (Prop. 1). Inference is a single tensorized forward pass (≈11 ms / 1K rows). On 17 ODDS benchmarks SPOT-FM obtains 0.915 average AUC-ROC and 0.752 AUC-PR, improving over FoMo-0D (Shen et al., 2024) on 12/17 datasets.