Google's RRSI Boosts AI Agent Generalization Without Test Memorization

Researchers from Google Cloud AI Research and partner universities have developed RRSI, a self-improvement method designed to enhance the generalization capabilities of AI agents. The approach avoids the common pitfall of test memorization, where agents optimize for specific benchmarks at the expense of broader performance. Testing on eight benchmarks with the Claude Opus 4.8 model revealed gains of up to 14.1 points on training tasks and up to 4.7 points on unseen benchmarks. The method also reduced execution tokens by approximately 30% without compromising performance.

The team found that traditional self-improvement cycles often lead agents to memorize test tasks, resulting in inflated training scores but diminished performance on new challenges. RRSI addresses this by limiting the scope of modifications and avoiding hardcoding of task-specific solutions. The system employs a critic to filter changes, ensuring only meaningful improvements are retained while obsolete components are removed.

RRSI operates by rewriting the harness that controls an agent's prompts, tools, and logic, shifting from broad overhauls to targeted adjustments over time. The method was also tested with the Gemini 3.5 Flash model, boosting the accuracy of the weaker Gemini 3.1 Flash Lite by 11.2 to 14.6 points. The research team has made the supporting code available on GitHub, emphasizing the reliability of self-improvement for enhancing agent capabilities.