RRSI: regularized agent self-improvement gains 4.7 points with 30% fewer tokens

HuaxiuYaoML · x · 2026-09-22

The paper "RRSI: Regularized Recursive Self-Improvement of Agent Harnesses" (Huaxiu Yao et al., Google team, arXiv Sep 21, 2026) adds regularization to automated agent harness evolution to prevent overfitting to task-specific quirks.

Results: up to 4.7-point gains on unseen benchmarks while using 30% fewer policy tokens than unregularized evolution. The authors position this as letting engineers building complex LLM workflows move beyond hand-crafted prompting and fragile DSPy pipelines toward automated harness evolution that resists memorization and controls inference cost.

Related event: Google's RRSI regularizes recursive self-improvement of agent harnesses(4 posts)→

Original post →

More from coding & agent

coding & agent channel →