Following up on the previous Nature Select team's empathy evaluation method, I also reviewed their MAPO training approach. The purpose of this training method is to enable the model to maintain a stable "empathy persona" during multi-turn interactions and provide relatively consistent responses.

In other words, the focus of this methodology is on the behavioral performance of the Agent itself, rather than understanding authentic psychological models, guiding human-computer interaction toward clearly defined "health/goodness."

Furthermore, both papers lack an operationalizable definition of "empathy." From the text, it can be understood roughly as "the ability to demonstrate understanding, support, reassurance, and non-aggressive responses in dialogue." This definition is clear but admittedly not sufficiently deep, and the conceptual ambiguity does not facilitate researchers in truly addressing the structural issues in Agent's human-computer relationship.

Original text: