Circuit Traces

The attribution graph method on an open model. Each case study below is a causal graph through Gemma 2 2B: which features caused which, from the prompt to the predicted token.

Method: Ameisen et al. and Lindsey et al. (2025), computed on Gemma Scope transcoders. A caveat that matters: these graphs describe a replacement model built from transcoder features, pruned for legibility. They are strong hypotheses about the real computation, to be confirmed with interventions, not proofs.