MindZero: Learning Online Mental Reasoning With Zero Annotations

2026-05-29

AI Summary

This work addresses the challenge of efficiently and robustly inferring human mental states in real-world scenarios where annotated data are scarce. To this end, the authors propose MindZero, a self-supervised reinforcement learning framework that internalizes model-based theory-of-mind reasoning into a single forward inference pass. By integrating a multimodal large language model with action-likelihood estimates derived from planner feedback, MindZero generates high-confidence psychological hypotheses without requiring ground-truth mental state annotations, enabling online mentalizing. Evaluated on grid-world and household tasks, MindZero significantly outperforms both pure language models and conventional model-driven approaches in terms of both accuracy and computational efficiency.

Abstract

Effective real-world assistance requires AI agents with robust Theory of Mind (ToM): inferring human mental states from their behavior. Despite recent advances, several key challenges remain, including (1) online inference with robust uncertainty updates over multiple hypotheses; (2) efficient reasoning suitable for real-time assistance; and (3) the lack of ground-truth mental state annotations in real-world domains. We address these challenges by introducing MindZero, a self-supervised reinforcement learning framework that trains multimodal large language models (MLLMs) for efficient and robust online mental reasoning. During training, the model is rewarded for generating mental state hypotheses that maximize the likelihood of observed actions estimated by a planner, similar to model-based ToM reasoning. This method thus eliminates the need for explicit mental state annotations. After training, MindZero internalizes model-based reasoning into fast single-pass inference. We evaluate MindZero against baselines across challenging mental reasoning and AI assistance tasks in gridworld and household domains. We found that LLMs alone are insufficient; model-based methods improve accuracy but are slow, costly, and limited by backbone MLLM capacity. In contrast, MindZero enhances MLLMs' intrinsic ToM ability and significantly outperforms model-based methods in both accuracy and efficiency, showing that mental reasoning can be effectively learned as a self-supervised skill.

Problem

• Theory of Mind
• online inference
• zero annotations
• mental reasoning
• real-time assistance

Innovation

• self-supervised reinforcement learning
• Theory of Mind
• online mental reasoning
• multimodal large language models
• zero annotations