Google Cloud AI Research, together with UNC‑Chapel Hill, Stanford and Washington University in St. Louis, introduced RRSI (Regularized Recursive Self‑Improvement). RRSI lets a language‑model agent rewrite its own harness—prompts, tools, memory, control flow and sub‑agents—without touching model weights. The improvement loop is regularized so gains transfer to unseen benchmarks.
Self‑improving harnesses usually overfit because they reuse the same evolve set each round, memorize task‑specific details, chase noise and accumulate unnecessary complexity. RRSI counters these failure modes with a set of concrete mechanisms.
On the proposal side, an annealed edit budget starts with bundled edits and narrows to single, attributable changes later. Every edit is logged with component, hypothesis, diff, score change and cost; the proposer reads this ledger to avoid retrying falsified ideas. When progress stalls inside the noise band, the budget shifts to untouched components for structured exploration.
On the selection side, a leakage critic strips task names, entities and benchmark‑specific logic before scoring. A noise‑adjusted floor requires gains to exceed variance measured on the base harness. A cost rule forces any extra inference tokens to be paid for by measurable gain. Components that stop yielding gains are marked for pruning. These rules map to classic regularizers: the edit budget to L0, pruning to L1 (Lasso) and the cost rule to L2 (Ridge).
Across eight benchmarks, all six held‑out splits improved. Using Gemini 3.5 Flash as the policy, Terminal‑Bench 2.1 rose from 64.6 to 78.7 and SWE‑bench Verified from 76.8 to 79.0. Token usage dropped from 3.80 M to 2.42 M per trial, a roughly 30‑36 % reduction.
Getting started is straightforward: clone the Apache 2.0 repository, install the dev dependencies, and run the provided scripts for a coding domain. New domains plug in via a single adapter module.
The core takeaway is that RRSI enables continuous harness evolution while keeping model weights frozen, delivering reliable out‑of‑distribution gains without the overfitting pitfalls of naïve self‑improvement loops.
#AI #MachineLearning #LLM #SelfImprovement #AIResearch #Tech