Self-Play Methods in Reinforcement Learning for Language Models

In this thesis we develop a series of practical algorithms for language model to self-train, by actively and strategically creating and controlling learning experiences themselves, enabling better model generalizability and training efficiency compared to traditional approaches.