This article details Direct Nash Optimization (DNO), a method designed to optimize LLMs using Nash equilibrium principles, addressing the challenges faced by traditional soft policy iteration. DNO replaces unstable and complex policy updates with a regression-based contrastive objective for stable batch training. The approach enjoys monotonic improvements and converges to the Nash equilibrium. A 7-billion-parameter LLM trained with DNO outperforms Mistral Large and earlier versions of GPT-4 on AlpacaEval 2.0. The paper highlights key design choices for developing iterative self-improving algorithms.




