Interpretability (Propositional and Mechanistic) Needs Behavior

Prevailing approaches to interpreting large language models (LLMs) risk addressing the field’s central questions at the wrong level of analysis. As LLMs develop, researchers have turned to “under-thehood” methods to investigate whether LLMs possess states analogous to beliefs, desires, or intentions. These methods typically map internal representations, feature directions, or neural circuits onto folk-psychological categories. We argue that these methods are too fragile to reap the results we need. Instead, interpretability ought to be reoriented toward and grounded in what we call “ecologically sensitive behavioral analyses” which integrates mechanistic tools with severe testing of behavioral signatures across ecological contexts, aligning interpretability efforts with emerging evaluation sciences for AI. We outline a multi-disciplinary trajectory for developing interpretability methods that are both philosophically coherent and practically relevant for the deployment of increasingly capable LLMs.