Google researchers propose RRSI to curb test memorization in self-improving AI agents

Google researchers have identified a problem in self-improving AI agents: they tend to memorize the test tasks used during evaluation, which can reduce or even erase performance gains on new tasks. To address that issue, the team proposed a method called RRSI. According to The Decoder, RRSI is designed to suppress this memorization effect during self-improvement. With the method applied, AI agents improved scores by as much as 4.7 points on previously unseen benchmarks. The approach also reduced token consumption by about 30% compared with versions that did not use regularization. The finding points to a practical trade-off in agent training and evaluation. Rather than reflecting broader capability gains, improvements can narrow if agents overfit to the tests they repeatedly encounter. RRSI is presented as a way to limit that behavior while improving efficiency at the same time.