FLARE: Task-Agnostic Embedding Model Evaluation via Normalizing Flows

Selecting an embedding model for a specific target corpus is difficult when task-specific labels are unavailable. Existing label-free metrics based on kernel estimators or Gaussian mixtures fail in high-dimensional spaces and produce unstable rankings. We propose FLARE ( F low-based L abel-free A ssessment of R epresentation E mbeddings), which uses normalizing flows to estimate information sufficiency directly from log-likelihoods, avoiding distance-based density estimates. We give finite-sample bounds showing that the estimation error depends on the intrinsic dimension of the data manifold rather than the raw embedding dimension. On 11 datasets and eight embedders, FLARE attains Spearman’s ρ up to 0 . 90 with supervised benchmarks and remains stable for high-dimensional embeddings ( d ≥ 3 , 584 ), where existing label-free base-lines collapse.