Sign InOpen Brain
arXivPaperNeeds Review

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

Mind2Dialogue trains assistants from simulated, evolving user states that are hidden at deployment. The approach could inform persistent agents, but its reported gains concern personalization rather than coding work.

arXiv · Sep 14, 2026
Open Source Open MarkdownOpen JSON
Source Summary

Mind2Dialogue uses a psychology-guided simulator to maintain personal traits while updating a shared mental state through conversation. An Oracle assistant sees that state, and **privileged distillation** trains deployed models to respond without direct access to it.

Practical Implication

Builders of persistent agents can treat inferred preferences and beliefs as uncertain training signals rather than facts stored in a profile. The evaluation joins personalization with theory-of-mind tasks across Qwen, Llama, and OLMo baselines.

Agent-Ready Context
Mind2Dialogue uses a psychology-guided simulator to maintain personal traits while updating a shared mental state through conversation. An Oracle assistant sees that state, and **privileged distillation** trains deployed models to respond without direct access to it.

Builders of persistent agents can treat inferred preferences and beliefs as uncertain training signals rather than facts stored in a profile. The evaluation joins personalization with theory-of-mind tasks across Qwen, Llama, and OLMo baselines.

The full corpus improved every reported personalization metric, including **26.6–40.9 percentage-point gains** in preference-following generation. The users and mental states are simulated, and the material gives no evidence for coding-agent workflows or real long-term deployments.
Connected Context · Feed7 Judgment

This adds a third memory design to the candidates: train a model from privileged, evolving representations of a user rather than inject stored context or consolidate repeated experience directly. It strengthens the case for treating user state as dynamic, but sharply narrows the evidence to simulated people and conversations; it does not establish that inferred beliefs should become durable deployment memory.

Context Map
model#agent-memory
Uncertainty
The full corpus improved every reported personalization metric, including **26.6–40.9 percentage-point gains** in preference-following generation. The users and mental states are simulated, and the material gives no evidence for coding-agent workflows or real long-term deployments.