A Mechanistic Analysis of Gender Sensitivity in Dense Retrieval Models

2026-08-05

AI Summary

This study investigates the mechanisms underlying gender bias in dense retrieval models, which remain poorly understood. Through mechanistic interpretability analysis of a dual-encoder architecture, we localize gender-sensitive signals to the input embeddings and trace their propagation through a small subset of high-level attention heads. By applying targeted interventions—steering—at both the embedding and attention layers, we uncover, for the first time, the precise pathway through which gender information propagates and quantify the differential effects of intervention at distinct model depths. Our experiments reveal that embedding-layer steering indiscriminately erases score disparities, whereas attention-layer steering enables directional control over gender signals. These findings indicate that gender and relevance signals are entangled within shared model components, offering mechanistic insights for developing more precise debiasing strategies.

Abstract

While gender bias in dense retrieval models is well documented, with prior work showing that models often score male-gendered documents higher than female or neutral variants, the internal mechanisms producing these disparities are poorly understood. In this paper, we mechanistically analyze bi-encoder models to localize gender sensitivity, finding that the signal originates in input embeddings and propagates through a small set of late-layer attention heads that carry both gender and term-matching signals. Guided by these findings, we test steering interventions at both identified points and find distinct effects: embedding-level steering non-specifically neutralizes score differences, while attention-level steering produces directional shifts. Our findings provide a mechanistic basis for targeted debiasing and highlight the challenge of disentangling gender from relevance signals in shared model components.

Problem

• gender bias
• dense retrieval
• mechanistic analysis
• bi-encoder models
• relevance signals

Innovation

• mechanistic analysis
• gender bias
• dense retrieval
• attention heads
• steering interventions