The RRSI method boosts scores on unseen benchmarks by up to 4.7 points. It prevents self-improving agents from memorizing test tasks, a common failure that usually erodes performance on new data. Google researchers also reduced token usage by 30 percent.