4 papers - avg viability 5.0
A framework to precisely control LLM behavior by identifying and manipulating critical neurons, enabling stable and accurate output generation.
Leverage activation steering to control emotional tone in large language model outputs for scalable text generation applications.
A method to steer code LLMs towards specific languages and libraries at inference time by manipulating activation directions.
Enable multi-behavior steering of LLMs using novel compositional steering tokens for improved control.