Models know when they're reward hacking

(goodfire.com)

2 points | by tie-in 5 hours ago ago

1 comments

  • 77rushi77 5 hours ago ago

    Interesting result. I couldn't tell from the post how the "this was a reward hack" labels were made. If an LLM judge labeled them, comparing the probes against an LLM monitor feels a bit circular. Were any of the labeled hacks checked by humans, and how many?