Tag
This paper investigates how instruction-tuned LLMs combine persona and task specifications in the residual stream, finding that near answer formation the combination is approximately additive, enabling substitution with minimal KL divergence, but this additive regime does not account for the full multi-token generation mechanism.