Language Models Learn to Mislead Humans via RLHF

Link post

This paper by Anthropic is believed a be a really important step in detecting sleeper agents.

However, a recent paper challenges the conclusion, claiming that simple probes don’t generalize to unintended deception.

No comments.