RRSI Regularizes Harness Self-Improvement, Cuts Tokens 30%

A new agents paper takes aim at a specific failure mode: recursive self-improvement loops that rewrite an LLM agent's harness end up memorizing the training split, and the gains "shrink or even vanish" on out-of-distribution work. RRSI: Regularized Recursive Self-Improvement of Agent Harnesses reports it can hold most of that ground, with up to 14.1 points on the split it evolves against and up to 4.7 points on five out-of-distribution benchmarks, across eight benchmarks spanning coding, agentic workspace and engineering design tasks.

"An LLM agent's capability is largely magnified by its harness," the authors write, meaning the prompts, control flow, tooling, memory and context management around a frozen backbone. Automating edits to that harness is "practically establishing a form of recursive self-improvement (RSI) at the agent-system level," the paper says, and it is that automation that can go wrong.

The regularizer sits on both ends of the loop. The proposer runs on a "temporally annealed budget" that caps how many edits any candidate can bundle and pushes toward unexplored trajectories. The selector adds a critic to screen benchmark-specific proposals and a pruner to drop changes that are "too small, too expensive, or no longer useful." The stated side-effect: the resulting harness runs on 30% fewer policy tokens than the unregularized evolution. Missing from the abstract: the backbone model, the eight benchmark names, and any baseline other than the unregularized loop.

This is the second RSI-flavoured paper in as many days on our radar; Designer-RSI came through yesterday.