Causal Localization of Evidence Effects in Transformers via CEA
October 28, 2026
Controlled evidence aggregation (CEA) identifies specific attention heads influencing model decisions by varying symbol counts and orders in prompts. Using Pythia-410M, the BUC method selects 1-12 heads whose ablation removes 65% of the count effect, providing a method to map symbolic inputs to specific transformer circuits.
HOW THIS AFFECTS YOU
●
researcherYou can use this method to more precisely attribute model behavior to specific attention heads during mechanistic interpretability tasks.