FlashREINFORCE: Critic-Free Single-Rollout Asynchronous RL for Agentic Language Models

Benchmark Model Rank Results
mathematical-reasoning-on-minerva-mathFlashREINFORCE (Qwen2.5-Math-1.5B)#40Accuracy: 22.8
mathematical-reasoning-on-olympiadbenchFlashREINFORCE (Qwen2.5-Math-1.5B)#46Accuracy: 34.5
mathematical-reasoning-on-math-500FlashREINFORCE (Qwen2.5-Math-1.5B)#114Accuracy: 69.2