A new arXiv paper, “From Encoding to Generation: Towards Empathetic Large Language Models,” proposes that empathy in AI should be understood as two separate capacities. The first, cognitive empathy, is about accurately representing another person's emotional state. The second, affective empathy, is about generating a suitable response to that state.
The authors argue that conflating these two abilities has limited progress in empathetic language models. By separating them, the paper suggests, researchers can more precisely target what an LLM lacks: either a poor internal model of the user's emotions or a poor mapping from that model to a response.
Because the source is only an abstract, the specific methods and results are not yet available. The contribution at this stage is conceptual: a clearer vocabulary for evaluating and building empathetic AI systems.