In other words: The influence of prompt variation on alignment evals

·LessWrong··

If alignment evaluations are phrased differently, does model behaviour change? In this post, we describe our attempt to answer this question. This project was completed as a 5-day project sprint during the Capstone Week of ARENA 8.0.TL;DR: The data we collected were noisy! Almost every eval, model, and prompt variation direction yielded inconsistent results, making it difficult to draw clear conclusions. Nonetheless, our data suggest that prompt rephrasing can have a measurable impact on alignme...

Read full article →

Related Articles

Private German rocket makes history, reaches orbit from European soil
bookmtn · Hacker News · 10h ago
LLMs as a Cognitive Virus
canjobear · Hacker News · 10h ago
Actively exploited sandbox RCE in all Chromium versions
negura · Hacker News · 1d ago
Formalizing Fermat's Last Theorem
jlebar · Hacker News · 1d ago
South African diamond mines are closing due to weak sales and lab-grown stones
bookofjoe · Hacker News · 9h ago