MindPower: Enabling Theory-of-Mind Reasoning in VLM-based Embodied Agents

Researchers led by Ruoxuan Zhang along with 9 other authors have developed MindPower, a novel framework that enables Theory-of-Mind reasoning in vision-language embodied agents, addressing the current limitation in AI's ability to understand mental states of others. The research paper, submitted to arXiv on November 28, 2025 and revised on February 24, 2026, has been accepted for presentation at CVPR 2026, highlighting its significance in the field of artificial intelligence. The team introduces this Robot-Centric framework to overcome the gap in existing benchmarks that focus solely on human mental states while ignoring the agent's own perspective, which hinders coherent decision-making and action generation in AI systems. MindPower represents a comprehensive approach to creating more sophisticated embodied agents by integrating four key components: Perception, Mental Reasoning, Decision Making, and Action. When processing multimodal inputs, the system first perceives both the environment and human states, then performs Theory-of-Mind reasoning to model both self and others, and finally generates decisions and actions guided by these inferred mental states. This holistic approach enables AI systems to consider not only what humans are doing but also what they might be thinking or intending to do, creating more natural and contextually appropriate interactions. The researchers also introduced Mind-Reward, a novel optimization objective that encourages Vision-Language Models (VLMs) to produce consistent Theory-of-Mind reasoning and behavior. This innovation addresses the challenge of aligning internal mental models with external actions, a critical aspect of creating believable and effective AI agents. In performance evaluations, the MindPower framework demonstrated significant improvements, outperforming even advanced models like GPT-4o by 12.77% in decision making and 12.49% in action generation tasks. These results underscore the potential of Theory-of-Mind capabilities to revolutionize how AI systems interact with humans in complex environments, paving the way for more intuitive and context-aware robotic assistants and virtual agents.