Princeton Trains 4B-Parameter LLM to 2700 Elo in Chess With Move Explanations

Princeton researchers trained a 4-billion-parameter LLM using reinforcement learning combined with explainability. The model reached a 2700 Elo chess rating and explains its moves, with the method transferable to robotics and other games.

• 4B-parameter model reached a 2700 Elo chess rating
• Training combines reinforcement learning with explainability
• Model explains moves grounded in chess principles
• Method transfers to robotics, games and computer use