Tag
This paper introduces a deployable per-instance, multi-layer activation steering technique for large language models, showing that optimal layer selection varies per input and can be predicted from the prompt embedding without gold labels at inference.