Persona vectors extracted from three domains are decorrelated from a general-expert vector, then an MLP gate trained in two stages decides how much of each persona to add to the transformer residual stream.
Advances in Neural Infor- mation Processing Systems36, 66044–66063 (2023)
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.CL 1years
2026 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
FOCUS: Decoupling Expert Personas in LLMs to Enhance Domain Expert Capabilities
Persona vectors extracted from three domains are decorrelated from a general-expert vector, then an MLP gate trained in two stages decides how much of each persona to add to the transformer residual stream.