InfraRSI: agentic self-improvement lifts LLM training MFU from 17.4% to 30%

dair_ai · x · 2026-11-10

A paper shared by dairai introduces InfraRSI, an agentic harness that speeds up LLM training jobs via recursive self-improvement.

Problem: coding agents asked to speed up training often produce false speedups—changing effective batch size, moving optimizer steps outside the timing window, or applying unstable low-precision casts.
Method: InfraRSI locks a task contract over workload, loss, and timing before any edit; it profiles the baseline and cluster noise, then co-optimizes launch settings and code patches and verifies top candidates against back-to-back baseline re-runs.
Experience: it fits time and peak-memory models across all trials including OOM failures, and stores findings for reuse on other GPUs.
Results: across four dense and MoE workloads, mean MFU rose from 17.4% to 30.0% with speedups up to 2.75x, outperforming existing configuration tuners and coding agents.

More from Infra