Google researchers find a way to keep self-improving AI agents from memorizing their tests

AI agents are a base model wrapped in a harness, the prompts, tools and memory that shape what the model sees. Google researchers found that letting a model rewrite its own harness makes it memorize test tasks, so gains vanish on new ones. Their RRSI method limits and shrinks the edits allowed each round, reporting up to 4.7 more points on unseen benchmarks with about 30 percent fewer tokens.

Summary written with AI by HowTechMoves, based on reporting by The Decoder.