Self-play training study reports gains in mathematical reasoning

Researchers report that training a Qwen3-4B-Base model on records from self-play board-game searches raised its mean score across six mathematics benchmarks from 24.1 to 36.6. The model's win rate on a held-out game also rose from 15% to 45%. Their preprint describes a method that turns game searches into training examples containing a preferred move, alternatives, possible replies and estimated outcomes. The reported transfer to unseen mathematics suggests a way to generate reasoning training data without human labeling. The results come from the tested model and benchmarks; the source does not establish the same gains for other models or workloads.