Modeling LLMs as holding latent beliefs yields high-accuracy behavior predictions

brwilder · x · 2026-09-10

Follow-up on the new paper: across several settings, treating LLMs as holding a latent degree of belief about an unknown state allows high-accuracy predictions of behavior, with accuracy rising sharply as general capability increases. Directly relevant to the recent anthropomorphization debate.

Related event: Paper: LLMs' latent beliefs can predict their behavior(2 posts)→

Original post →

More from Research

Research channel →