Tag
This paper investigates whether large language models have a localized causal mechanism for handling the animacy concept, using circuit discovery on minimal pairs; they find an animacy circuit that is distributed and only partially generalizes.