simpo

A deep dive into DPO (Rafailov et al. 2023) and its successors — how they reformulate RLHF as a classification problem, the math behind the implicit reward, and where each variant wins or loses against PPO-based pipelines.