Paired-Model Experiment: Misaligned-Reward CoTs Are Indistinguishable From Aligned Ones
johnhewtt · x · 2026-10-20
Part of the same hewittlab thread: the researchers trained pairs of models with the same intent (pass all tests), one on the true reward and one on a misaligned reward (pass some, fail some). In thinking rollouts, both models only talk about the prompt intent — making the misaligned model unmonitorable.
Related event: Study: Reward Misspecification Makes Chain-of-Thought Unmonitorable(9 posts)→
More from Safety
- ByteDance, Alibaba, Tencent order 3M+ Nvidia RTX PRO 5500 cards worth $35B, Reuters reports — SumitGup · 2026-10-21
- CVSS 10.0 RCE vulnerability disclosed in PHP core, patches only for PHP 8 — cyb3rops · 2026-10-21
- Antigen launches Assets: a security agent that maps your infrastructure and triages new CVEs — alishbaimran_ · 2026-10-21
- GitHub HQ Hosts Open Source AI + Policy Talk with Mozilla Co-founder Mitchell Baker — kevinsxu · 2026-10-21
- Researcher Finds Two Windows 0-Days, First in 76 Minutes with a Frontier Model — moyix · 2026-10-21
- Stanford's Noah Goodman shares interactive AI safety concept map — stanfordnlp · 2026-10-21