Stabilizing bias removal in language models through group-relative training

Original: BiasGRPO: Stabilizing Bias Mitigation in High-Variance Reward Landscapes via Group-Relative Policy Optimization

Source: arxiv.org ↗