Agentic RL 后训练资源怎么分？港中文、恒生大学提出 Libra，吞吐最高提升 3 倍

TL;DR - Libra dynamically reallocates GPUs between rollout and training during agentic RL post-training, addressing workload drift and highly variable trajectory lengths. On 48 NVIDIA A800 GPUs, it delivered up to 3× baseline throughput and reached target rewards up to 2.5× faster.

• A global planner jointly optimizes training parallelism, heterogeneous rollout instances, and GPU allocation.
• An elastic hybrid pool shifts resources without rebuilding core training communication groups.
• C-MLFQ routes trajectories using causal signals from tool responses, achieving 91.1% routing accuracy on Search-R1.
• Tests across Search-R1, DAPO-Math-17K, and R2E-Gym showed the highest throughput among evaluated systems at comparable final rewards.