Anthropic 研究：对齐审计智能体自动发现问题成功率达 42%
In March we published a paper on alignment audits: teams of humans were tasked to find the problems in model we trained to be misaligned.
Anthropic 发布新研究，开发了三个 AI 智能体来自主完成对齐审计任务，测试中成功发现了隐藏目标、构建了安全评估并暴露出令人担忧的行为。Jan Leike 引用该研究指出，今年 3 月发表的论文中由人类团队负责找出被训练成失准模型的问题，如今智能体已能自动完成这一任务，成功率为 42%。
推荐理由
AI 评分 54
对比人类团队与自动智能体在模型对齐审计中的表现，给出智能体自主发现问题的比例。
主题
查看原帖New Anthropic research: Building and evaluating alignment auditing agents. We developed three AI agents to autonomously complete alignment auditing tasks. In testing, our agents successfully uncovered hidden goals, built safety evaluations, and surfaced concerning behaviors. https://t.co/HMQhMaA4v0
正文
当前显示原文，中文全文译文尚未就绪。
In March we published a paper on alignment audits: teams of humans were tasked to find the problems in model we trained to be misaligned.
Now we have agents that can do it automatically 42% of the time.