@diatkinson
New COLM paper: Identifying Introspection From the Inside
When an LLM tells us about its decisions, does it 𝘬𝘯𝘰𝘸 what drives its choices—or is it guessing?
In our setting, we find that faithful models decide and report with the same layers. Unfaithful ones don't. 🧵View quoted post on X
View quoted post on X
This is why I think fraud in science needs to be dealt with a lot more aggresively. The US should be investing a lot more into the US Office of Research Integrity (@HHS_ORI) so it can hire a team of field agents who carry out investigations, similar to what the FBI or SEC does.View quoted post on X
Sometimes I wish our language had tenses or other markers for levels of epistemic confidence. A one-syllable way to say:View quoted post on X
"I believe X because I have mathematical proof"
"I'm underqualified on this topic, but my feeling is X"
"My impression from observations so far is X"
etc
