Tag
This paper tests whether decodable empathy directions in LLMs can reliably shift automated empathy scores, finding that affective facet control is partial and cognitive steering is inconsistent, highlighting that detection does not imply control.