GameSmith

GameSmith: Scaling Zero-Sum Self-Play with Synthesised Games

Post-training large language models (LLMs) on two-player zero-sum games provides a verifiable reward, and the gains have been shown to transfer to general reasoning.