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

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

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

Zixuan Wang^1* Yufan Zhou^2* Jinzhou Tang^1* Xinle Yu¹ Chengjun Wu¹ Lyumanshan Ye¹ Zhaoxiang Feng¹ Letian Peng¹ Adyasha Patra¹ Fan Bai⁵ Enze Ma³ Zhengding Hu¹ Jianyang Gu⁴ Zhao Wang¹ Yufei Ding¹ Jingbo Shang¹ Tianmin Shu⁵ Zhiting Hu¹ Zhen Wang^1‡

¹ UC San Diego² KU Leuven³ University of Illinois Chicago⁴ The Ohio State University⁵ Johns Hopkins University

* Equal contribution ‡ Corresponding author

arXiv Code Dataset

Method

PDF ↗

M2D-Sim shares an evolving user state with the Oracle to generate dialogue and QA supervision for a student. — M2D-Sim shares an evolving user state with the Oracle. Dialogue students learn from the persona and history, with the state withheld.

Dataset taxonomy

Preview of the scenario distribution. — Distributions in the paper’s analysis subset. Behavior modes are adapted from TUNA; fallback modes are excluded. Select a chart for the full figure and labels.

Preview of the persona-specialization distribution. — Distributions in the paper’s analysis subset. Behavior modes are adapted from TUNA; fallback modes are excluded. Select a chart for the full figure and labels.

Results

PDF ↗

Qwen2.5-7B-Instruct, the strongest same-backbone baseline per task, and M2D-Chat on personalization and theory of mind. — Qwen2.5-7B-Instruct comparisons against the strongest evaluated same-backbone baseline per task (paper Tables 1–2). AutoToM scores higher on ToMi. BigToM requires both true- and false-belief variants to be correct.

Citation

@misc{wang2026mind2dialoguetraininghumanawarelanguage,
      title={Mind2Dialogue: Training Human-Aware Language Models by Simulating User Mental States},
      author={Zixuan Wang and Yufan Zhou and Jinzhou Tang and Xinle Yu and Chengjun Wu and Lyumanshan Ye and Zhaoxiang Feng and Letian Peng and Adyasha Patra and Fan Bai and Enze Ma and Zhengding Hu and Jianyang Gu and Zhao Wang and Yufei Ding and Jingbo Shang and Tianmin Shu and Zhiting Hu and Zhen Wang},
      year={2026},
      eprint={2609.15972},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.15972},
}