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

Training human-aware language models is hindered by a supervision gap: existing datasets lack responses grounded in the unspoken beliefs, goals, and changing circumstances that define real human interactions. The authors investigate whether simulating these latent user states can provide the privileged supervision necessary to train assistants that better understand and support their users.

The Mind2Dialogue framework consists of three components:

• M2D-SIM: A psychology-guided simulator that generates multi-turn conversations. It maintains a structured, evolving user state (beliefs, goals, emotions) that influences both the simulated user's behavior and the assistant's response.
• Oracle Supervision: An Oracle assistant is given direct access to the user's latent state during training data generation. This allows the Oracle to produce responses that address the user's underlying intentions rather than just their surface-level utterances.
• Privileged Distillation: Student models are fine-tuned on these Oracle-generated dialogues. Crucially, the student model is trained only on the observable dialogue history, with the latent state withheld, ensuring the model learns to infer and act on user states without needing them as explicit inputs at deployment.

Training on the Mind2Dialogue corpus consistently improves personalization across three different model families (Qwen, Llama, and OLMo). On Qwen2.5-7B-Instruct, the model achieved significant gains in preference-following generation (up to 40.9 percentage points) and outperformed several task-specific baselines. Furthermore, the training improved belief and action reasoning (theory of mind) for Qwen and Llama, demonstrating that the framework fosters capabilities beyond simple personalized assistance. However, the authors note that these reasoning gains are not universal, as the OLMo model showed mixed results on mental-state reasoning tasks.