Causal Localization of Evidence Effects in Transformers

Overview

arXiv:2610.36524v1 Announce Type: new Abstract: Can controlled symbolic inputs reveal attention heads that influence a language model's decisions from natural language evidence? We introduce controlled evidence aggregation (CEA), which varies the counts and order of two symbols in otherwise fixed prompts and measures changes in a frozen Transformer's preference between two answer tokens. The preference is measured by their logit difference; the prompts prescribe no correct answer. Balanced utili

SpinGraph analysis pending — check back after processing.