Gemma models recall facts via a three‑phase circuit that leans on the residual stream
Activation‑patching on Gemma‑2B and Gemma‑12B‑IT shows that factual recall proceeds in three stages—encoding, routing, and read‑out—mostly via the residual stream. The study maps these phases across transformer layers, revealing a consistent circuit that scales with model size. Code and data are released for reproducibility.