MindZero: Learning Online Mental Reasoning With Zero Annotations

MindZero introduces a self-supervised reinforcement learning framework that trains multimodal large language models to perform robust Theory of Mind reasoning without requiring annotated mental state data. The approach combines model-based planning with neural scaling, achieving superior accuracy and efficiency compared to traditional model-based methods and LLMs alone.

MindZero addresses a fundamental challenge in AI alignment and human-computer interaction: understanding human mental states from observable behavior without expensive manual annotation. The framework leverages a clever training signal where the model learns to generate mental state hypotheses that best explain observed actions through an internal planning mechanism. This self-supervised approach sidesteps the annotation bottleneck that has historically limited Theory of Mind research to small datasets.

The research builds on two converging trends: the growing recognition that large language models lack robust reasoning about human psychology, and the effectiveness of reinforcement learning for instilling behavioral understanding into neural networks. Prior work demonstrated that model-based methods could achieve better accuracy than LLMs, but these approaches sacrifice computational efficiency and scalability. MindZero synthesizes these insights by training the neural network to internalize planning-based reasoning as a learned skill, eliminating the runtime cost while preserving accuracy gains.

For the AI industry, this work signals that foundation model capabilities in human-centric reasoning are not fundamental limitations but rather learnable skills achievable through appropriate training objectives. This has immediate applications in AI assistants, robotics, and personalization systems that must infer user intent. The self-supervised training paradigm also democratizes Theory of Mind research by removing reliance on scarce annotated datasets, potentially accelerating development across academia and industry.

The critical next steps involve evaluating MindZero on real-world behavioral data beyond simulated gridworlds and household domains, and determining whether learned reasoning transfers across diverse contexts and user populations.

• →MindZero eliminates the need for manual mental state annotations by using action-likelihood maximization as a self-supervised training signal.
• →The framework internalizes model-based reasoning into fast single-pass inference, outperforming both pure LLM approaches and slower model-based baselines.
• →Self-supervised Theory of Mind training represents a scalable pathway to improving AI systems' human-centered reasoning capabilities.
• →The approach demonstrates that computational efficiency and interpretability need not be sacrificed when scaling neural networks for social reasoning.
• →Real-world deployment requires validation beyond gridworld simulations to prove robustness across diverse human behaviors and contexts.