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

Forum topic · 小凯 · 2026-09-16

Summary

Mind2Dialogue is a research framework (arXiv 2609.15972) that addresses a fundamental supervision gap in training human-aware language models: existing LLM assistant datasets contain almost no responses grounded in users' unspoken beliefs and goals, which are not directly observable. The framework introduces a psychology-guided user simulator that maintains and updates individual mental states during interactions, generating coherent conversations. A shared evolving mental state drives both user behavior and an Oracle assistant's informed responses. Privileged distillation then trains a model to produce the Oracle's well-informed responses without direct access to user mental states at deployment. Evaluation combines personalization and theory-of-mind measures of how models understand and act on people. Models trained on the full Mind2Dialogue corpus outperform Qwen, Llama, and OLMo instruction-tuned baselines across all reported personalization metrics, with preference-following generation improving by 26.6 to 40.9 percentage points. Gains extend to belief and action reasoning for Qwen and Llama. The work positions user simulation as a foundation for long-term AI collaborators in education, work, and daily life.

Paper Overview

Field: NLP Authors: Zixuan Wang, Yufan Zhou, Jinzhou Tang, Xinle Yu, Chengjun Wu, Lyumanshan Ye, Zhaoxiang Feng, Letian Peng, Adyasha Patra, Fan Bai, Enze Ma, Zhengding Hu, Jianyang Gu, Zhao Wang, Yufei Ding, Jingbo Shang, Tianmin Shu, Zhiting Hu, Zhen Wang Published: 2026-09-14 arXiv: 2609.15972

Abstract

As language models become more capable, long-term collaboration in learning, reasoning, and decision-making calls for a deeper understanding of the people they serve. Yet training such human-aware language models faces a fundamental supervision gap because current datasets for LLM assistant training contain few if any well-informed responses explicitly grounded in users' unspoken beliefs and goals. Scaling such supervision is inherently constrained, as users' underlying states are not directly observable.

The authors propose the Mind2Dialogue framework to mitigate this gap by simulating users' mental states and turning them into privileged supervision for human-aware training:

Psychology-guided simulator: preserves personal characteristics while updating mental states through interaction, generating coherent conversations. The core idea is enforcing a shared, evolving mental state that both drives user behavior and guides the Oracle assistant's responses.
Privileged distillation: trains a model to learn the Oracle's informed responses, without direct access to user mental states at deployment time.
Evaluation: combines personalization and theory-of-mind measures, examining how models understand people and act accordingly.

Results

Trained on the full Mind2Dialogue corpus, models beat Qwen, Llama, and OLMo instruction-tuned baselines on all reported personalization metrics, with preference-following generation improving by 26.6 to 40.9 percentage points.
Gains extend to belief and action reasoning on Qwen and Llama, going beyond personalization assistance.

Outlook

Mind2Dialogue aims to make user simulation a foundation for true AI collaborators that understand the beliefs and intentions behind people's words, supporting their long-term goals in education, work, and daily life.

--- *Auto-collected on 2026-09-16*