Attribution Graph
A graph tracing which internal features caused a model’s output, used to reverse-engineer its reasoning.
An attribution graph is an interpretability artifact: nodes are internal features the model computed, edges are the causal influence between them, and the graph traces how the input flowed through features to produce a specific output. It turns “the model said X” into a mechanistic account of which concepts fired and drove the answer.
Worked example: an interpretability technique that traces how a model reached an output by mapping which internal features fed into which — building a graph of the computation so you can see the ‘reasoning’ path rather than just the answer. Gotcha: the graph is a reconstruction, not a literal wiring diagram — it approximates the model with interpretable components and can miss or misattribute steps, so it is evidence about a mechanism, not proof; it is powerful for forming hypotheses about model behavior that you then test with interventions.