Google 研究提出 RRSI,防止自改进智能体记忆测试任务
中文摘要
Google提出RRSI,通过收缩编辑预算和严格审查,防止自改进智能体在改写harness时记忆测试任务。
English Summary
Google's RRSI prevents self-improving agents from memorizing test tasks via shrinking edit budgets and strict critic reviews during harness rewriting.
原文节选
Google 研究人员提出 RRSI(Regularized Recursive Self-Improvement of Agent Harnesses),用逐步收缩的编辑预算和严格 critic 审查,抑制语言模型反复改写智能体 harness 时对测试任务的记忆。