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

·LessWrong··

Epistemic status: Confident in the direction and approximate size. The effects are large and all 7 monitors moved the same way. Physics only, natural (non-adversarial) errors, one fixed group of monitors.This is a follow up to The Answer is Not the Argument which covered the observational results. This post covers an intervention. TLDR: We gave 7 chain-of-thought monitors step-numbered solutions to Humanity's Last Exam physics questions and asked them to find the first erroneous step. We then he...

Read full article →

Related Articles

Every AI model has the same "controversial" take on humans
rakannimer · Hacker News · 5h ago
Jailbreaking the Switch 2
ledoge · Hacker News · 13h ago
The recently proposed o(n log n) integer multiplication algorithm [pdf]
craigcitro · Hacker News · 19h ago
Security Disclosure
passenger · Hacker News · 1d ago
The 1.5°C goal of the Paris Agreement is no longer viable
velo_aprx · Hacker News · 1d ago