Circuit Tracing
Following the internal steps a model took on one specific prompt to see how it reached its answer.
Definition
Circuit tracing is a method for working out how a model produced a particular answer by following the chain of internal steps that led there. Researchers swap parts of the network for more readable stand-ins, then trace which of those readable pieces influenced which others on the way to the output, producing a map called an attribution graph. Unlike asking a model to explain itself — which only shows what it says about its reasoning — circuit tracing looks at what actually happened inside. Anthropic released open-source circuit-tracing tools in 2025, and the approach has become one of the main practical techniques in mechanistic interpretability.