Google unveils rrsi: AI framework that enables self-improvement without overfitting

Google Cloud AI Research, in collaboration with UNC-Chapel Hill, Stanford, and Washington University in St. Louis, has released RRSI (Regularized Recursive Self-Improvement), a framework that allows large language models (LLMs) to rewrite their own harness components, including prompts, tools, memory, and control flow, without altering model weights. The system constrains the improvement loop to ensure gains are consistent across benchmarks the agent hasn't optimized against. RRSI is deployable as a research framework, using Apache 2.0 licensing and requiring Python 3.10+. It supports any LiteLLM model string, with defaults assuming Claude Opus 4.8 on Vertex AI. The framework improves performance on benchmarks like Terminal-Bench 2.1, where scores rose from 64.6 to 78.7, and SWE-bench Verified, which increased from 76.8 to 79.0. The research team also highlights a 30% reduction in policy tokens per trial, with the project page noting a 36% improvement. The framework addresses three failure modes: benchmark-specific fitting, noise chasing, and complexity accumulation. By regularizing the search process, RRSI maintains edit budgets akin to L0, pruning like Lasso (L1), and cost rules similar to Ridge (L2). It outperforms other methods in out-of-distribution performance, with an OOD average more than 1 point above the baseline.