Teaching Models to Infer User Mental States
Mind2Dialogue moves away from simple instruction tuning by training models to track evolving user mental states.
Dialogue models typically map input tokens directly to output responses using statistical correlations observed in training data. This approach struggles when a user's intent is obscured by misunderstood constraints or shifting goals. Mind2Dialogue incorporates theory-of-mind labels into the training process. These labels document the causal relationship between an inferred mental state and the request, bridging the gap between raw input and specific user intent.
The Mechanism of Privileged Distillation
The training process operates in two stages. First, a simulator generates synthetic dialogue history mapped to latent states like the user's current goal or knowledge level. In the second stage, a student model undergoes privileged distillation, where it is trained on the Oracle model's well-informed responses. The student receives both the dialogue history and the latent state labels as input during training.
By conditioning the student's output on these labels, the training forces the model to minimize the discrepancy between its predicted response and the target response while provided with causal context. Because the input includes these labels, the attention mechanism assigns higher weights to tokens reflecting the latent state. This process embeds the causal dependencies directly into the model's weight distributions. During inference, the explicit labels are removed, but the model preserves these associations through its internal hidden layer representations. The model effectively performs state tracking by treating raw tokens as sufficient triggers to activate the latent causal pathways it learned during training. These reasoning patterns are unique because the simulator provides explicit counterfactuals and logic chains that standard instruction-tuning datasets often lack.
| Feature | Standard Dialogue Model | Mind2Dialogue |
|---|---|---|
| Input Processing | Raw tokens | Tokens and state labels |
| Training Signal | Text-to-text mapping | Causal state-to-intent |
| Inference Requirement | None | None |
Evaluating Intent Inference
To see why this works, consider a user who asks "Why is my file not uploading?" after previously stating they have a poor network connection. A standard model might suggest restarting the software or checking permissions. A model trained via Mind2Dialogue is conditioned on the latent state of the conversation—the recorded latency constraint—to prioritize troubleshooting the network.
Performance is measured via preference-following generation. This metric measures whether the model's output aligns with the user's specific documented constraints and latent goals. An improvement here signifies that the model is no longer guessing based on general internet trends but is responding to the user's unique mental model. Empirical results show gains ranging between 26.6 and 40.9 percentage points in preference-following generation metrics, depending on the base architecture and the density of the training data.
What remains unclear is how the model handles long-tail constraints that fall outside the simulator's original training parameters. While the model learns to associate history with state-dependent logic, the limit of this generalization when the user introduces novel goals not represented by the simulator is not yet known. Building on this requires a clear understanding of the simulator's coverage, as the student model is essentially a mirror of the synthetic reasoning it was fed.