Mapping First-Token Language Identity Circuits in Multilingual Models
September 28, 2026
Circuit analysis using Edge Attribution Patching identifies the directed acyclic graphs responsible for language broadcasting in models like Pythia-2.8B and BLOOM-560M. The study maps how multilingual models resolve target response languages during early generation steps.
HOW THIS AFFECTS YOU
●
researcherYou can gain mechanistic insights into how multilingual models decide on a target language at the first token.