Steering Vectors in LLMs Work by Hijacking Value Computations, Not Attention Patterns — and 90-99% of the Vector Is Dead Weight

Steering Vectors in LLMs Work by Hijacking Value Computations, Not Attention Patterns — and 90-99% of the Vector Is Dead Weight | BedrockNews