Framework for Unifying Omnimodal Embedding Models

Embedding research is converging toward common principles via contrastive learning and Transformer architectures. A new framework evaluates this convergence through three capabilities: coverage, comparability, and compositionality.