Association ≠ \neq Similarity: Learning Corpus-Specific Associations for Multi-Hop Retrieval

Jason Dury · 2026

Read Association ≠ \neq Similarity: Learning Corpus-Specific Associations for Multi-Hop Retrieval by Jason Dury (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-20850.txt.

Description

Dury presents Association-Augmented Retrieval, a lightweight re-ranking method that trains an MLP to learn relationships between texts using contrastive learning on co-occurrence tags. On HotpotQA, AAR lifts passage Recall@5 from about 83 percent to about 92 percent without tuning, with gains on hard questions. An inductive model trained on training links and tested on unseen links shows no significant gain, hinting the method captures corpus-only links rather than portable patterns. The method adds about 4 ms per query and needs under two minutes of single-GPU training.

Catalog record

CREATORJason Dury
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE36830
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.20850

Browse the full library · About the Tower of Babel