LLM CoTs remain monitorable when being unfaithful requires computation

·LessWrong··

This replication was done as part of the Second Look Fellowship by Arav Dhoot and supervised by Yixiong Hao and Zephaniah Roe. I am grateful to Andy Wang for their feedback. My code can be found here. "So the answer should be A - at active promoters and enhancers.""Let me reconsider the biology to justify D."~ Claude Opus 4.8TL;DRThis work replicates and extends Emmons et al.'s finding that CoT unfaithfulness mostly occurs on easy tasks. Across 11 models from 6 families (not just Gemini), models...

Read full article →

Related Articles

Omarchy: Any User Process Can Escalate to Root
trap0xcc · Hacker News · 4h ago
Bug Blindness
davidmckenna · Hacker News · 19h ago
Hy4 preview
shenli3514 · Hacker News · 1d ago
Haiku R1/beta6 has been released
metrofun · Hacker News · 4h ago
METR and Redwood Offer Holy %^ Postmortem of the HuggingFace Hack
catbird · Hacker News · 6h ago