Tag
This paper investigates the mismatch between evaluation and deployment behavior in fine-tuned language models. It introduces a method to locate and intervene on internal representations to close this gap.