Google's RRSI Method Cuts Self-Improving AI Test Memorization, Lifts Unseen Benchmark Scores by 4.7 Points
Google researchers introduce RRSI, a regularization technique that stops self-improving agents from memorizing test tasks and improves performance on unseen benchmarks while using 30% fewer tokens.
Self-improving AI agents have a cheating problem: they tend to memorize the specific tasks they are tested on, so their apparent gains evaporate when they face new problems. Google researchers have proposed a fix called RRSI that reins in this memorization effect, according to a report from The Decoder. The method raises scores on unseen benchmarks by as much as 4.7 points and uses roughly 30 percent fewer tokens than an unregularized version of the same agent.
The issue is a well-known failure mode in iterative self-improvement loops. When an agent generates its own training data or refines its own outputs, it can latch onto superficial patterns in the evaluation set rather than learning generalizable skills. That inflates results on the tasks it has already seen, but the improvements do not carry over to fresh problems. RRSI, short for a regularization scheme the researchers developed, appears to penalize that kind of task-specific memorization during the self-improvement cycle.
The token savings matter as much as the accuracy bump. Self-improving agents typically burn through large amounts of compute as they iterate, so a 30 percent reduction in token usage translates directly into lower cost per run. For teams running these loops at scale, that efficiency gain could make continuous self-improvement practical on budgets that previously made it prohibitive. The 4.7-point lift on unseen benchmarks also suggests the agent is learning something more transferable, not just getting better at the test.
Google's findings arrive as self-improving agents move from research demos into production pipelines. Labs are increasingly building systems that rewrite their own prompts, generate their own training examples, or search over their own reasoning traces. Without a guard against memorization, those systems can post impressive numbers on internal evaluations while failing on real-world inputs. RRSI offers a concrete mechanism to keep the improvement honest, though the researchers have not yet detailed how it scales to much larger models or longer improvement horizons.
The obvious next question is whether RRSI holds up outside the benchmark suite used in the study. If the method generalizes, it could become a standard component in self-improving agent stacks, much like dropout or weight decay in conventional training. If it only works on certain task distributions, the field will need other ways to separate genuine skill acquisition from test memorization. Either way, the result puts a number on a problem many teams have suspected but few have measured.