Researchers behind a new paper on Hugging Face have introduced Mind2Dialogue, a framework for training language models that are aware of the people they assist. The work addresses what the authors call a fundamental supervision gap: existing datasets for training LLM assistants rarely contain well-informed responses explicitly grounded in a user's unspoken beliefs and goals.
The challenge is structural, since users' underlying mental states are not directly observable. Mind2Dialogue works around this by simulating those states and treating them as privileged supervision. A psychology-guided simulator preserves personal characteristics while updating mental states through interaction, generating coherent conversations driven by a shared, evolving mental state that shapes user behavior and guides an Oracle assistant's replies.
The second stage, privileged distillation, trains models on the Oracle's well-informed responses so they can assist users at deployment without direct access to their mental states. The authors also propose evaluating human-aware learning by combining personalization with theory of mind, testing both how models understand people and how they act on that understanding.
According to the paper, training on the full Mind2Dialogue corpus improves every reported personalization metric over corresponding Qwen, Llama, and OLMo instruction-tuned baselines, including gains of 26.6 to 40.9 percentage points in preference-following generation. The authors report that gains extend to belief and action reasoning on Qwen and Llama, beyond personalized assistance.
The team frames the work as a step toward AI collaborators that grasp the beliefs and intentions behind people's words and support long-term goals in education, work, and daily life. The dataset and code have been released for the community.