trace-tournaments

Reasoning Arena improves reinforcement learning with verifiable rewards by using trace tournaments and Bradley-Terry models to generate meaningful gradients from non-diverse reward groups, resulting in faster training and better reasoning performance.