METR：Notes（网页）·· 16 小时前AI 评分55
METR 发布逐动作阻断式 Agent 监控器的实现与评估笔记
Implementing and Evaluating a Basic Per-Action Monitor for Safer Evals September 27, 2026 Reilly Haskins, Rif A. Saurous, Nate Rush, Neev Parikh, and Beth Barnes lay out the claims that would need to hold for their per-action blocking monitor to be effective, the evidence for each claim, and the gaps that remain. Read more
AI 导读
METR 研究人员开发并部署了一个逐动作监控器，用 LLM judge 在 Agent 每个工具调用执行前打分，3/10 及以上暂停评估并交人工审查，仅针对真实世界伤害和监控破坏行为。
来源：METR：Notes（网页） · metr.org