ReFlow preprint reports up to 2.06× higher AI training throughput

Researchers report that ReFlow, a framework for training large language models with reinforcement learning, improves training throughput by up to 2.06× over leading synchronous and asynchronous systems. Experiments covered single-turn reasoning and multi-turn agent tasks, with comparable or better model performance. The framework reuses recent training trajectories instead of consuming each once, spreading the cost of generating them across multiple uses. It also adjusts computing resources and generation strategies as workloads change. The preprint's reported gains apply to the representative tasks tested; the maximum improvement was not reported as uniform across workloads.