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

Mind2Dialogue uses a psychology-guided simulator of user mental states to produce privileged supervision, improving every personalization metric for Qwen, Llama, and OLMo assistants, including 26.6-40.9 percentage point gains in preference-following.

Mind2Dialogue is a framework for training human-aware language models. A psychology-guided simulator preserves users' personal characteristics while updating their mental states through interaction, generating coherent conversations and responses from an Oracle assistant grounded in those evolving states. Privileged distillation then trains models on the Oracle's well-informed responses, so that deployed assistants can help users without direct access to their mental states. Training on the full corpus improves every reported personalization metric over Qwen, Llama, and OLMo instruction-tuned baselines, including 26.6 to 40.9 percentage point gains in preference-following generation, and evaluation combines personalization with theory-of-mind belief and action reasoning. The story is covered by two sources — Hugging Face daily papers (2026-09-13) and an arXiv listing under cs.AI, cs.LG, and cs.CL (2026-09-14) — which report identical facts with no disagreements.

• A psychology-guided simulator preserves personal characteristics while updating user mental states through interaction, driving coherent conversations.
• An Oracle assistant, grounded in the simulated mental states, provides well-informed responses used as training targets.
• Privileged distillation trains assistants on the Oracle's responses; at deployment, models lack direct access to users' mental states.
• Training on the full corpus improves every reported personalization metric over Qwen, Llama, and OLMo instruction-tuned baselines.
• Preference-following generation improves by 26.6 to 40.9 percentage points over the baselines.
• Evaluation combines personalization with theory-of-mind belief and action reasoning.
• Sources: Hugging Face daily papers (2026-09-13) and arXiv cs.AI/cs.LG/cs.CL (2026-09-14); both reports agree on all facts.