Anomaly detection turns to diversity: VAS tops 148 embedding benchmarks without hyperparameter tuning
Synopsis
The Vendi Anomaly Score (VAS) redefines an anomaly as how much dataset diversity drops when a sample is removed, scoring samples through changes in the eigenvalue spectrum of the Vendi score rather than through density estimation or space partitioning. Across 148 ADBENCH image and text embedding benchmarks, VAS holds a better average rank than compared methods including ADERH, DN2 and SLAD, with no validation-set tuning; in controlled ImageNet contamination experiments it remains strongest as contamination rate and dataset scale grow.
Interpretation
Anomaly detection gains a criterion that relies on neither density nor partitioning: whether a sample is anomalous depends on how much overall diversity falls when it is removed from the dataset. Earlier unsupervised methods relied on isolation difficulty, local density at a fixed neighborhood scale, or distance from a fitted distribution, making them sensitive to neighborhood size, partitioning choices and distributional assumptions. Across the 148 ADBENCH embedding benchmark instances, VAS has an average AUROC rank of 2.35, against 3.80 for the best competing method GMD and 4.05 for DN2, with a Wilcoxon signed-rank test reporting a significant difference.
Scoring needs one global eigendecomposition followed by batched quadratic-form computation for all samples, so cost grows linearly with dataset size. Recomputing the leave-one-out Vendi score for every sample would require N eigendecompositions, cubic in scale, and is impractical for large embedding sets. With a first-order approximation replacing per-sample recomputation, the approximation preserves anomaly rankings with an average Spearman correlation above 0.98 across the 148 benchmarks, and AUROC changes by at most about 1.16 percentage points.
The Vendi score order is set automatically from the data's principal eigenvalue, using small orders for concentrated embeddings and larger orders for isotropic ones, removing neighborhood selection and tuning. Prior work using Vendi scores typically fixed a single order or swept a range to draw diversity profiles, while unsupervised anomaly detection offers no validation set for tuning. On ImageNet with DINOv3 ViT-B/16 embeddings, under both a fixed normal-class set with varying anomaly rate and a fixed anomaly rate with growing dataset scale, VAS achieves the strongest performance among the evaluated methods.
Perspective
The demonstrated scope is unsupervised anomaly detection on embedding datasets: ADBENCH image and text embeddings, and ImageNet embeddings extracted with DINOv3 ViT-B/16; the method uses only embeddings and cosine similarity, not labels, clean reference sets or model confidence scores.
VAS depends on the quality of the embedding space and the chosen similarity function; if anomalies are indistinguishable from normal samples under the embedding or cosine similarity, VAS cannot reliably detect them. The scalar score does not by itself explain the source of anomalousness, and the adaptive rule for selecting the order is heuristic and may not be optimal for all data geometries.
