Empathy Is Steerable but Multi-Axial: Mechanism Geometry and Persona Effects in LLMs
EMNLP 2026 Main ConferenceFirst Author
Abstract
This work investigates the activation-level structure of supportive empathy in instruction-tuned LLMs. Using the EPITOME framework and contrastive activation steering, we examine the controllability and representational geometry of emotional reactions, interpretations, and explorations, as well as their interaction with persona-induced activation shifts.
