Representation-Level Audit Reveals Bias Mitigation Success in Foundation Models
A new study analyzes how bias mitigation impacts the embedding spaces of BERT and Llama2, showing reduced gender-occupation disparities. The findings highlight measurable improvements in model fairness through representational analysis.