Skip to content

Translation in progress Some entries still include text in the original language.

NVIDIA · News

Oct 13 · Tuesday

NVIDIA · News

Latest publications and developments
Oct 4Sun
  1. The DecoderThe Decoder · 48

    Google 提出 RRSI 防止自改进 Agent 记忆测试任务

    Google 研究者提出 RRSI(Regularized Recursive Self-Improvement of Agent Harnesses),在让语言模型反复改写 Agent harness 的自我优化循环中加入约束,以避免 Agent 记忆测试任务。论文指出,自我优化会让 Agent 记住有限的测试任务,训练分数上升但新任务收益缩小甚至消失。原因包括:搜索记住只适配单一基准的模式、偏好偶然高分的候选,以及堆叠无助于能力的复杂度。

You have reached the end