InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure

InfraRSI. Harnessing Recursive Self-Improvement for Training Infrastructure. By InfraRSI Team and ten others, from Peking University. Published on the eighth of November.

Recursive Self-Improvement, or RSI, has sped up the synthesis of isolated algorithms and kernels. Distributed training of large language models is a different terrain. Its launch settings, operator kernels and communication schedules are tightly coupled, and a single edit can run into an out-of-memory boundary, a compiler conflict, measurement jitter, or silent numerical drift. An optimizer working here has to survive all four, and must not quietly win by changing what is being measured.

InfraRSI is an agentic harness that takes open-ended commands from a person and turns them into verified training optimizations. Its organising device is a five-stage workflow, with guardrails placed so that each gain it reports is one that was checked against a fixed baseline.

The stages run in order. Init locks down the editable scope and a fixed task contract covering workload, loss, and timing invariants, the guard against reward hacking. Doctor profiles the baseline's phases and calibrates the noise of the cluster. Run co-optimizes launch knobs and code patches, then re-verifies the top candidates against the baseline and logs what it found to an experience bank shared across hardware. Report and apply audit the source paths, track what failed trials cost, and commit the changes that were promoted.

The search itself avoids stepping greedily from one incumbent. It starts from diverse configurations seeded by experience, and fits global models of execution time and of censored peak memory across every trial, the out-of-memory ones included. Counting the failures matters: a trial that ran out of memory still tells the model something about where the boundary lies. With those models it does three things. It prices the memory cost of each individual setting. It revises LLM priors that turn out to be miscalibrated, using what live runs actually show. And it pairs memory-hungry optimizations with memory savers and with verified code patches, so that a setting too expensive on its own can still be used when something else pays for it.

The test bed is four workloads, single-node and multi-node, covering dense and Mixture-of-Experts language models. InfraRSI reaches a speedup of up to two point seven five times. Mean Model FLOPs Utilization rises from seventeen point four percent to thirty percent. Put another way, the share of the hardware's compute doing useful model work goes from under a fifth to nearly a third. InfraRSI outperforms both existing tuners and coding agents, and the experience it gathers on one kind of GPU transfers to others.

Self-improvement appears here at two levels. Inside a run, priors are revised online as trials come back. Across clusters, experience is distilled and carried forward into the bank. Those two together are the two-level loop the name refers to, and its stated purpose is the one the whole harness serves: making training infrastructure efficient enough to push foundation model scaling further.