Researchers discover that bias mitigation in BERT and Llama2 creates measurable geometric changes in embedding spaces, reducing gender-occupation stereotypes.

Study analyzes how debiasing techniques reshape internal representations in both encoder-only (BERT) and decoder-only (Llama2) foundation models

Findings show bias mitigation significantly reduces gender-occupation disparities, producing more neutral and balanced embeddings

Representational shifts are consistent across different model architectures, suggesting fairness improvements have interpretable geometric patterns

Embedding analysis validated as an effective tool for auditing and measuring the success of debiasing methods in foundation models