EMPA: Evaluating Persona-Aligned Empathy as a Process

Researchers introduce EMPA, a new framework for evaluating persona-aligned empathy in LLM-based dialogue agents by treating empathetic responses as sustained processes rather than isolated interactions. The system uses controllable scenarios and multi-agent testing to assess long-term empathetic behavior in AI systems.

• →EMPA evaluates AI empathy as a continuous process rather than individual responses to better assess real-world performance.
• →The framework uses psychologically grounded scenarios coupled with multi-agent sandbox testing to expose AI behavioral patterns.
• →Current evaluation methods struggle with latent user states and sparse feedback in empathetic AI interactions.
• →The system scores AI trajectories based on directional alignment, cumulative impact, and stability in psychological space.
• →The framework extends beyond empathy evaluation to other AI agent settings with weak feedback mechanisms.