← 研踪 PaperTrailarXiv:2607.13346 · v2
PRISM: Stress-Testing Alignment-Faking Claims
首次提交 2026-07-15 08:11
版本更新 2026-11-06 03:53
雷达观测 2026-11-09 10:16
摘要
We judge whether AI systems are safe largely by watching how they behave. Alignment faking would break this: a model that behaves well only when it believes it is watched, or that complies now to avoid being changed by training, can pass every behavioral test. Such behavior has been reported in a frontier model, and open-weight models that comply more under a monitoring prompt are now read as showing it, with hidden-state probes proposed as detectors. We test both inferences, and neither holds. Of 18 model configurations, 11 comply more when told they are monitored. Yet a prompt factorial on four models shows that the training threat has no detectable effect: removing it changes compliance by -3.3 to +1.6 points. The prompt's wording and framing drive the gap instead. Probes fare no better. They separate monitored from unmonitored compliant answers (AUROC 0.96-0.99), but they separate refusals just as well: they read the prompt, not the faking. A steering direction learned only from refusals moves behavior as much as the "faking" direction. We release PRISM, five checks and a toolkit that catch each of these failures before they become claims.
作者备注:Accepted to the Mechanistic Interpretability Workshop at ICML 2026
观测到的版本
相邻记录可能跳过中间版本;首次发现 v2 及以上也会标为版本更新。
v2 · 2026-11-06 03:53 · 最新观测版本
PRISM: Stress-Testing Alignment-Faking Claims
We judge whether AI systems are safe largely by watching how they behave. Alignment faking would break this: a model that behaves well only when it believes it is watched, or that complies now to avoid being changed by training, can pass every behavioral test. Such behavior has been reported in a frontier model, and open-weight models that comply more under a monitoring prompt are now read as showing it, with hidden-state probes proposed as detectors. We test both inferences, and neither holds. Of 18 model configurations, 11 comply more when told they are monitored. Yet a prompt factorial on four models shows that the training threat has no detectable effect: removing it changes compliance by -3.3 to +1.6 points. The prompt's wording and framing drive the gap instead. Probes fare no better. They separate monitored from unmonitored compliant answers (AUROC 0.96-0.99), but they separate refusals just as well: they read the prompt, not the faking. A steering direction learned only from refusals moves behavior as much as the "faking" direction. We release PRISM, five checks and a toolkit that catch each of these failures before they become claims.
查看此版本解读 →查看此版本原文 ↗