LLMs are getting pretty good at “understanding” human preferences. This is indeed part of my hope. But goal misgeneralization is still an issue for RL systems based on LLMs, and monolithic LLMs…

Posters: davidad. davidad

3 likes recorded by the source corpus.