谷歌RRSI限制AI代理记忆测试任务
谷歌研究人员开发RRSI，防止AI代理记忆测试任务，提高在未见基准上的性能。这影响AI代理开发。
google researchers propose regularized recursive self-improvement (rrsi) to address the issue of ai agents memorizing test tasks. rrsI optimizes the harness, a framework controlling the agent's behavior, by limiting edits and using a strict critic to evaluate proposals. tested on eight benchmarks with claude opus 4.8, rrsI improves performance on unseen tasks by up to 4.7 points and reduces token usage by 30%. this development affects the design of self-improving ai agents.