InfraRSI Team (Yuchuan Tian, Haocheng Mei, Ye Yuan, Chao Xu, Boxun Li, Guohao Dai, Kai Han; TokenRhythm) · Agentic Training Optimiser

InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure

InfraRSI is an agentic harness that turns human commands into verified distributed LLM training optimisations through a five-stage workflow, including a fixed task contract to prevent reward hacking and a cross-hardware experience bank. It fits global execution-time and peak-memory models that include out-of-memory trials. Across four single- and multi-node dense and Mixture-of-Experts workloads, the paper reports a speedup of up to 2.75x and mean Model FLOPs Utilisation rising from 17.4% to 30.0%, outperforming existing tuners and coding agents.