Theory of mind inspired mental state reasoning with large language models for empathetic response generation
Year: 2026
Type: article
Abstract: As conversational AI expands into emotionally sensitive domains such as mental health support and counseling, empathetic response generation has become a key research challenge. Existing empathetic dialogue systems either reflect the user’s surface level emotions or, when relying on external commonsense knowledge bases, capture only generic patterns. We propose a Theory of Mind inspired framework in which a large language model first infers the speaker’s five mental states of Belief, Desire, Emotion, Intent, and Knowledge from the multi turn dialogue history and then uses them as conditioning signals for empathetic response generation. A preliminary study on a theory of mind benchmark indicates that the LLMs employed perform reasonably on mental state inference in both multiple choice and free form settings. On an empathetic dialogue dataset, injecting the self generated inferences improved response quality under both automatic metrics and LLM as judge evaluation. In the LoRA fine-tuned setting, the ToM condition performed comparably to COMET, an external commonsense knowledge baseline, with higher ROUGE-L point estimates for all four backbones and a significant advantage for one. Shuffled and corrupted descriptions consistently degraded response quality, whereas the ToM condition and a context-grounded factual-summary control performed similarly, indicating that contextual grounding is the primary source of the benefit. These findings suggest that speaker grounded mental state reasoning can be an effective strategy for empathetic dialogue systems.
Source: Scientific Reports
Authors: Hayoung Lee, Seonghee Lee
Language: English
Topic: Speech and dialogue systems
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
Award: G9001391857
OpenAlex ID: W7234420833
