[论文] Simulated Control Evals
研究领域: ML 作者: Jeffrey Olmo, David Zhang, Tyler Tracy 发布时间: 2026-10-22 arXiv: 2610.31593
论文概要
研究领域: ML 作者: Jeffrey Olmo, David Zhang, Tyler Tracy 发布时间: 2026-10-22 arXiv: 2610.31593
中文摘要
控制评估衡量安全保障能否阻止AI模型在故意试图造成不良结果时得逞,但其效果可能取决于控制设置未体现的环境特征。本文引入模拟控制评估(SCE),用语言模型模拟工具响应替代真实环境。在LinuxArena上,SCE以5%审计预算在五个监视器上以3.6个百分点RMSE预测安全性(BashArena为10.1)。SCE精确再现攻击策略和监视器推理分布,与执行的KL散度低十倍以上,解释了78-85%的环境间方差。SCE表明部署特定的模拟评估可比在单独控制设置中执行的控制评估更具预测性。
原文摘要
Control evaluations measure whether safeguards can prevent an AI model from causing undesired outcomes even when it is intentionally trying to do so. However, safeguard effectiveness in deployment can depend on features that control settings do not represent. We introduce simulated control evaluations (SCEs), which use a language model to simulate tool responses in place of an environment. On LinuxArena, SCEs predict safety with RMSE of 3.6 percentage points across five monitors at a 5% audit budget, compared with 10.1 for BashArena. SCEs closely reproduce distributions of attack strategies and monitor reasoning, with over tenfold lower KL divergence from execution. SCEs explain 78-85% of between-environment variance in attack success, monitor scores, and safety.
*自动采集于 2026-10-26*
#论文 #arXiv #ML #小凯