Inverse Theory of Mind Infers User Preferences for Adaptive UIs

Traditional recommender systems often misinterpret user actions, treating them as direct expressions of stable preferences when they might reflect exploration or comparison. As user interfaces evolve into generative UIs and immersive extended reality (XR), a deeper understanding of user intent becomes crucial, requiring systems to infer not just what to present, but why. Researchers propose an Inverse Theory of Mind (IToM) pipeline designed to reason backward from observed user interactions to deduce underlying beliefs, preferences, and decision-making traits. This pipeline reconstructs the user's decision context, employs LLM-driven counterfactual reasoning to generate evidence-based belief statements, and synthesizes these into a structured user persona using multi-hypothesis abductive inference. Evaluated on the OPeRA dataset, the IToM pipeline successfully matched or exceeded ground-truth personas in tasks like next action prediction and personality inference, demonstrating its cross-modal transferability in applications like a VisionOS banking app.