Google Cloud AI Research and university collaborators developed RRSI, a method that prevents self-improving AI agents from memorizing their test tasks. The approach caps and gradually shrinks edit budgets and uses a critic to reject benchmark-specific tricks. Tested on eight benchmarks with a frozen Claude Opus 4.8 model, RRSI gained up to 14.1 points on training tasks and 4.7 points on unseen benchmarks, while using about 30 percent fewer tokens.
No score is assigned. Sources and their independence are shown in the citation chain below.