Training for Compositional Sensitivity Reduces Dense Retrieval Generalization

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Ralev, Radoslav, Baral, Aditeya, Zhechev, Iliya, Agarwal, Jen, Rajamohan, Srijith
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866914484919992320
author Ralev, Radoslav
Baral, Aditeya
Zhechev, Iliya
Agarwal, Jen
Rajamohan, Srijith
author_facet Ralev, Radoslav
Baral, Aditeya
Zhechev, Iliya
Agarwal, Jen
Rajamohan, Srijith
contents 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.
format Preprint
id arxiv_https___arxiv_org_abs_2604_16351
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Training for Compositional Sensitivity Reduces Dense Retrieval Generalization
Ralev, Radoslav
Baral, Aditeya
Zhechev, Iliya
Agarwal, Jen
Rajamohan, Srijith
Information Retrieval
Artificial Intelligence
Computation and Language
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.
title Training for Compositional Sensitivity Reduces Dense Retrieval Generalization
topic Information Retrieval
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2604.16351