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.
Cite
@article{arxiv.2604.16351,
title = {Training for Compositional Sensitivity Reduces Dense Retrieval Generalization},
author = {Radoslav Ralev and Aditeya Baral and Iliya Zhechev and Jen Agarwal and Srijith Rajamohan},
journal= {arXiv preprint arXiv:2604.16351},
year = {2026}
}