Training for Compositional Sensitivity Reduces Dense Retrieval Generalization

ArXi:2604.16351v1 Announce Type: cross Dense retrieval compresses texts into single embeddings ranked by cosine similarity. While efficient for recall, this interface is brittle for identity-level matching: minimal compositional edits (negation, role swaps) flip meaning yet retain high similarity. Motivated by geometric results for unit-sphere cosine spaces (Kang, 2025), we test this retrieval-composition tension in text-only retrieval.