InfraRSI: Optimizing Training Infrastructure with Recursive Self-Improvement

InfraRSI is an agent framework that turns human instructions into verified distributed LLM training optimizations and uses a fixed task contract to prevent reward hacking. On four dense and MoE workloads, it achieves up to 2.75x speedup, raising average MFU from 17.4% to 30.0%, outperforming existing tuners and coding agents. It corrects priors online and distills experience across clusters, forming a two-level recursive self-improvement loop.