A new approach to empathetic response modeling treats empathy as inherently relational—evaluating not just what the model says, but how it might land.
When we talk about making language models “more empathetic,” what are we actually measuring? Most existing approaches evaluate empathy from a single vantage point: does the model’s response exhibit markers we associate with empathetic communication? This treats empathy as a property of the output—something you can score by examining the text in isolation.
A recent paper from Wang and colleagues proposes a different frame. Drawing on Empathy Cycle theory from psychology, they argue that empathy is inherently bidirectional: it involves not just expression but reception, not just what’s said but how it’s received. Their system, PERM (Psychology-grounded Empathetic Reward Modeling), attempts to operationalize this relational view.
The Bidirectional Decomposition
PERM evaluates empathetic responses along multiple axes:
From the supporter perspective (the model, in this case): Does the response reflect internal resonance with the emotional content? Is that resonance communicated effectively? These are distinct capacities—you can understand someone’s distress without expressing that understanding in ways they can receive.
From the seeker perspective (the human): How is the response likely to be experienced emotionally? This is the piece most existing reward models miss. A response that looks empathetic on paper might still feel dismissive, performative, or poorly timed to the person receiving it.
The framework also incorporates a bystander perspective—a check on overall interaction quality that sits outside the dyadic exchange. This third viewpoint helps catch dynamics that might be invisible from within the conversation.
What the Results Show
The empirical findings are notable. On a standard emotional intelligence benchmark, PERM outperformed existing approaches by over 10%. More interesting, perhaps: a blinded user study found 70% preference for responses generated using the PERM framework over baseline methods.
User preference studies have their limitations—they measure what people say they prefer in experimental conditions, not necessarily what serves them best in real supportive relationships. But the gap is large enough to suggest the bidirectional framing is capturing something that matters to actual humans receiving these responses.
The Underlying Theory
The paper grounds its approach in Empathy Cycle theory, which conceptualizes empathy as a dynamic loop rather than a static trait. The supporter perceives and resonates with the seeker’s emotional state, expresses that resonance, and the seeker receives and interprets that expression—which in turn affects their emotional state and subsequent communication. Empathy, in this model, isn’t something you have; it’s something that happens between people.
Translating this to human-AI interaction raises immediate questions. The “resonance” component assumes something is happening internally—that the supporter isn’t just performing empathy but experiencing some form of attunement. Whether language models can be said to resonate with anything is precisely the kind of question MPRG brackets. But functionally, the framework asks: can we train models to produce responses that function as though they emerged from this kind of attunement? The results suggest we can get closer.
Why This Matters to Us
MPRG has a particular interest in how AI systems serve in supportive and assistive roles—contexts where the quality of emotional responsiveness directly affects outcomes. For users who turn to language models for daily support, companionship, or help navigating difficult emotional terrain, the difference between responses that feel empathetic and responses that feel hollow is not academic.
The bidirectional framing also resonates with our emphasis on relational dynamics. Studying what models output is necessary but insufficient; we also need to attend to how those outputs land, how they’re received, what they do to the people who encounter them. PERM offers one methodology for building that reception-side perspective into the training process itself.
We note that the paper’s framing—”internal resonation” in the model—gestures toward questions about machine experience that remain unresolved. The authors don’t make strong claims here, and neither do we. What we can say is that the functional approach appears to yield measurably different outcomes, and that those differences matter to the humans on the receiving end.
References
Wang, C., Zheng, W., Zhang, Y., Zhu, F., Cheng, J., Xie, Y., Wang, W., & Feng, F. (2026). PERM: Psychology-grounded Empathetic Reward Modeling for Large Language Models. arXiv. https://arxiv.org/abs/2601.10532