New study reveals where gender bias hides in dense retrieval models

Dense retrieval models are known to rank male-gendered documents higher than female or neutral variants, but the internal cause has remained a black box. In a new arXiv paper, Catherine Chen, Maarten de Rijke, and Carsten Eickhoff open that box. By mechanistically analyzing bi-encoder architectures, they traced the gender-sensitive signal back to two specific locations: the input embeddings and a small set of late-layer attention heads. These heads simultaneously carry both gender and term-matching signals, explaining why simple bias fixes often fail—the two functions are entangled in the same components.

To test practical implications, the team applied steering interventions at each identified point. Embedding-level steering non-specifically neutralized score differences across the board, essentially washing out relevance sensitivity. Attention-level steering, by contrast, produced directional shifts—able to push scores toward or away from gender-specific phrasing, but not cleanly eliminate bias. This distinction is crucial for real-world search and recommendation systems. The findings give developers a mechanistic basis for targeted debiasing, but also warn that without decoupling gender from relevance in shared attention heads, any fix risks degrading retrieval quality. The paper is available on arXiv and offers a new roadmap for building fairer, more transparent retrieval models.