RSI-Opt preprint reports lower training error with evolving optimizer agents

A preprint reports that RSI-Opt, a system that revises the instructions and tools guiding AI training agents, reduced mean relative L₂ error on the Burgers PDE task by approximately 2.7-fold with Opus 5.5 and 8.8-fold with GPT-6.1 against a fixed-prompt agent baseline. A meta-agent uses evidence from earlier training runs to revise a prompt, procedural playbook, tested optimizer code and decision cadence. An archive retains evaluated policies for further evolution. Across 5 generations on three training tasks, the agents developed reusable procedures and code and ran CPU proxy experiments. Continued adaptation improved outcomes on unseen tasks, although transfer across PDE tasks varied by model and task.