Training For Compositional Sensitivity Reduces Dense Retrieval Generalization

Abstract

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 et al., 2025), we test this retrieval-composition tension in text-only retrieval. Across four dual-encoder backbones, adding structure-targeted negatives consistently reduces zero-shot NanoBEIR retrieval (8-9% mean nDCG@10 drop on small backbones; up to 40% on medium ones), while only partially improving pooled-space separation. Treating pooled cosine as a recall interface, we then benchmark verifiers scoring token--token cosine maps. MaxSim (late interaction) excels at reranking but fails to reject structural near-misses, whereas a small Transformer over similarity maps reliably separates near-misses under end-to-end training.

Stats

• citations0
• S2 citations—
• github stars0
• HF likes0
• heat score0.00
• arxiv keyralev2026training

Related papers

• BERM: Training The Balanced And Extractable Representation For Matching To Improve Generalization Ability Of Dense Retrieval (2023)5.84
• Dense Retrievers Can Fail On Simple Queries: Revealing The Granularity Dilemma Of Embeddings (2025)2.86
• Maximal Matching Matters: Preventing Representation Collapse For Robust Cross-modal Retrieval (2025)2.26
• Back To Basics: A Simple Recipe For Improving Out-of-domain Retrieval In Dense Encoders (2023)0.00
• Pylate: Flexible Training And Retrieval For Late Interaction Models (2025)3.58
• SCOT: Self-supervised Contrastive Pretraining For Zero-shot Compositional Retrieval (2025)0.00
• From Mapping To Composing: A Two-stage Framework For Zero-shot Composed Image Retrieval (2025)0.00
• Data-efficient Generalization For Zero-shot Composed Image Retrieval (2025)2.26