The Crowded Embedding Space: A Mean-Field Mechanism for Emergent Marginalization in Retrieval-Augmented Agents

Shwan Ashrafi, Dan Roth · 2026

Read The Crowded Embedding Space: A Mean-Field Mechanism for Emergent Marginalization in Retrieval-Augmented Agents by Shwan Ashrafi, Dan Roth (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-2606-28343.txt.

Description

Shwan Ashrafi and Dan Roth show how retrieval systems can quietly push minority interests aside, with no one intending it. Retrieval-augmented agents fetch documents from a shared embedding space, where each document sits at a point. Serving a majority taste can pack that neighborhood so densely that a similar minority taste gets crowded out of the top results.

Ashrafi and Roth build a mean-field framework, which tracks the crowd as a whole rather than document by document. In a fixed space they find a phase transition: past a critical majority density, minority performance collapses. Letting embeddings update to maximize accuracy makes things worse, since the dynamics drive the system toward serving only the majority. Relevance-chasing alone, they prove, manufactures this exclusion.

Catalog record

CREATORShwan Ashrafi, Dan Roth
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE71 KB
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2606.28343

Browse the full library · About the Tower of Babel