Embedding models now underpin many HR recommendation systems: resumes, job descriptions, and queries are converted into dense vectors, and candidate matches come from nearest-neighbor retrieval in that space. Because these models influence hiring decisions, any social bias they carry can have real-world consequences.
A recent arXiv preprint argues that the source of such bias is not only the training data but also the fine-tuning strategy. The paper, which presents a statistically rigorous audit of gender and stereotypical bias, suggests that choices made during model adaptation can either amplify or mitigate unfair outcomes.
The contribution is primarily methodological: a framework for auditing bias in embedding-based HR systems rather than a single set of findings. For practitioners, the implication is that fairness should be considered at the fine-tuning stage, not just when curating data.