Mechanistic Interpretability of LLM Behavior: A Survey

Mechanistic interpretability seeks to explain how internal computations produce large language model behaviors, yet findings remain fragmented across methods and behavioral domains. We survey 229 papers across five behavior families: epistemic, linguistic, reasoning, safety, and social behaviors. We introduce a framework distinguishing five levels of mechanistic evidence, from decodability to generalization. Our cross-behavior synthesis reveals recurring mechanisms, distinctions between representation and causal control, and limitations of existing explanations. We conclude by identifying gaps in understanding computational roles, behavioral interactions, and generalization to interactive and agentic settings.