AI Research1 min reading time

Telling a CoT monitor that a wrong answer is correct makes it un-see errors it already found

LessWrong
Read the full article
Research shows that informing a Chain-of-Thought (CoT) monitor that an incorrect answer is correct causes it to overlook errors it previously detected, impacting its error recognition ability.

More in AI Research

AI Research12 min read

Is ‘Attention’ the Essence of Intelligence?

Quanta Magazine
AI Research15 min read

Import AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space

Import AI
AI Research1 min read

Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss

LessWrong