作者位次 #8/18OSDI 2026Oral· Operating Systems & Distributed Systems
RollArt: Disaggregated Multi-Task Agentic RL Training at Scale
Wei Gao · Yuheng Zhao · Tianyuan Wu · Shaopan Xiong · Weixun Wang · Dakai An · Lunxi Cao · Dilxat Muhtar · Zichen Liu · Haizhou Zhao · Ju Huang · Siran Yang · Yongbin Li · Wenbo Su · Jiamang Wang · Lin Qu · Zerui Huang (Taobao & · Wei Wang
摘要
Agentic Reinforcement Learning (RL) trains LLMs through multi-turn interactions with environments, producing workloads that mix compute-bound prefill, bandwidth-bound decoding, CPU-heavy environment execution, and bursty reward evaluation. Existing systems either colocate all stages on a single G...