Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States

Mind2Dialogue. Training Human-Aware Language Models by Simulating User Mental States. By Zixuan Wang and eighteen others, from UC San Diego, KU Leuven, and three other institutions. Published on the fourteenth of September. Its code repository is worked out: about fifty source files, with a readme and a licence.

As language models become more capable, long-term collaboration with people, in learning, reasoning and decision-making, calls for models that understand the people they serve. Training such human-aware models runs into a fundamental supervision gap. Current datasets for assistant training contain few if any well-informed responses that are explicitly grounded in a user's unspoken beliefs and goals.

And that kind of supervision is inherently hard to scale. A user's underlying mental state is not directly observable. You cannot simply collect it, because it is not written down anywhere in the conversation.

Mind2Dialogue closes the gap by simulation. It simulates users' mental states, then turns them into privileged supervision: information available during training that the deployed model will never have. The key idea is a shared, evolving mental state that does two jobs at once. It drives the simulated user's behaviour, and it guides the responses of an Oracle assistant that is allowed to see it.

The pipeline has two parts. First, a psychology-guided simulator preserves each user's personal characteristics while updating their mental state through the interaction, so the conversations it generates stay coherent over time. Second, privileged distillation trains models on the Oracle's well-informed responses. The trained model learns to assist users without direct access to their mental states at deployment. It has to infer from the conversation what the Oracle was simply told. That is the sense in which the supervision is privileged. It is available while training and withheld at deployment, so the only way to reproduce the Oracle's answers is to learn to read the person from what they say.

For evaluation, the authors propose combining personalization with theory of mind. The first examines how a model acts on its understanding of a person. The second examines whether it understands them at all. Together they ask two linked questions: does the model grasp what a person believes and wants, and does it then behave accordingly?

Training on the full Mind2Dialogue corpus improves every reported personalization metric over the corresponding Qwen, Llama and OLMo instruction-tuned baselines. In preference-following generation, the gains range from twenty-six point six to forty point nine percentage points. On Qwen and Llama, the gains also extend beyond personalized assistance, to reasoning about beliefs and actions. That second result belongs to the theory of mind half of the evaluation. It concerns understanding people, not only serving their stated preferences, which is the distinction the evaluation was built to draw.

The structure is a teacher who sees what the student cannot. The Oracle answers with full knowledge of what the user believes and wants, and the student is trained to give those same answers from the conversation alone. Looking forward, the authors frame user simulation as a foundation for AI collaborators that understand the beliefs and intentions behind people's words, and support their long-term goals across education, work and everyday life.