A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
Self-Play Preference Optimization (SPO) is a minimalist reinforcement learning from human feedback algorithm that avoids training a separate reward model by using self-play.
Unknown
This paper proposes a minimalist offline RL approach that avoids out-of-distribution value overestimation by constraining the policy to the data support, achieving strong performance with simple modifications.
Zihe Liu, Jiashun Liu, Yancheng He, et al.
This paper systematically reviews RL techniques for LLM reasoning, identifies a minimalist combination of two techniques that outperforms GRPO and DAPO.