RLHF for Finer Alignment with Gemma 3

This 10-minute presentation from Google explores techniques for Reinforcement Learning from Human Feedback (RLHF) developed specifically for Gemma 3 models. Learn about the challenges of using reward models as proxies for human preferences, including the inevitable problem of reward hacking during prolonged training. Discover the innovative approaches implemented by the Gemma team to mitigate these issues and enable extended training periods, resulting in better alignment of language models for human interaction. Speaker Louis Rouillard explains how these techniques create finer alignment between AI systems and human preferences, ultimately producing more helpful and appropriate AI responses.