NVIDIA Open-Sources FlashREINFORCE: Half Rollout Cost, Better Accuracy

NVIDIA published in September 2026 the open method FlashREINFORCE, integrated into the Molt framework. It trains AI agents with half the rollouts of GRPO while achieving comparable or better accuracy on math and tool-use benchmarks.

• Reinforcement learning cost reduced by 50%
• Accuracy no worse than GRPO on math and tool-use tasks
• Method integrated into Molt framework
• Potentially doubles experiment throughput on GPU clusters