Reinforcement Learning (RL) for LLMs

Video Description

Lecture on reinforcement learning (RL) fine-tuning of large language models (LLMs). Even though we are in the RL era for training LLMs, this didn't start with DeepSeek R1, or even ChatGPT. The talk takes a deep dive through the history of RL training of LLMs, including my own early work on RL from human feedback (RLHF). Then we discuss more recent techniques to achieve personalized RLHF, and the future of RL for LLMs, including using multi-agent RL for adversarial red-teaming.