Initiative Vector Controls Turn-Taking in Full-Duplex Spoken Models

A single late-layer residual-stream direction, the initiative vector, allows for zero-shot control of speech ratios in spoken language models. Scaling this vector shifts model turn-taking from 6-12% to 42-61% speech ratio using only a short reference recording.