Group Relative Policy Optimization (GRPO)

31 Oct 2024

🚧 Work in progress…

This article will cover Group Relative Policy Optimization (GRPO), a recent advancement in policy optimization methods for training language models with human preferences.

Topics to cover:

• Motivation: improving upon PPO and DPO
• GRPO algorithm and key innovations
• Group-based relative preference modeling
• Empirical results and comparisons