circuit discovery

Once you accept that behaviors are implemented by circuits, the next problem is finding them without hand-tracing every wire. Circuit discovery automates the search for the minimal subgraph of heads, neurons, and connections that is responsible for a specific task — say, the components needed to identify the correct object in a sentence. The output is a small wiring diagram, carved out of a network with billions of connections, that still reproduces the behavior on its own.

The standard approach is to start from the full computation graph and prune. You repeatedly ask, via activation patching, whether removing or rerouting an edge hurts performance on the task; edges that do not matter are cut, leaving the causal skeleton. Automated methods like ACDC, edge attribution patching, and SAE-based circuit tracing scale this pruning so it can run over many components instead of a researcher's intuition.

Discovery is progress, not proof. A recovered circuit is only as good as the task and metric used to define it, circuits found for one prompt distribution may not generalize, and pruning can miss backup components that quietly substitute for ablated ones. The found subgraph is a hypothesis about the mechanism that still needs faithfulness checks.