Modeling the Mental World for Embodied AI: A Comprehensive Review

TL;DR

This review defines Mental World Models (MWMs) as the social-cognitive counterpart to traditional Physical World Models (PWMs) for embodied AI, and clarifies their differences and connections through predictive coding and POMDP formalisms. It surveys two core ToM reasoning paradigms—Prompting and Model-Based Inference—covering representative methods, tradeoffs, and neuro-symbolic integrations, and catalogs 26 ToM benchmarks to trace the evolution from static, text-based tasks to multimodal, dynamic interactions. The paper argues that a unified MWM framework, supported by neuro-symbolic fusion and online learning, is essential for robust social intelligence in embodied agents, and outlines practical, ethical considerations for real-world deployment. Collectively, this work provides a roadmap for advancing social cognition in AI through integrated theory, scalable benchmarks, and principled reasoning architectures with attention to safety and fairness.

Abstract

As the application of Embodied AI Agents in avatars, wearable devices, and robotic systems continues to deepen, their core research challenges have gradually shifted from physical environment interaction to the accurate understanding of social interactions. Traditional physical world models (PWM) focus on quantifiable physical attributes such as space and motion, failing to meet the needs of social intelligence modeling. In contrast, the Mental World Model (MWM), as a structured representation of humans' internal mental states, has become the critical cognitive foundation for embodied agents to achieve natural human-machine collaboration and dynamic social adaptation. However, current MWM research faces significant bottlenecks: such as fragmented conceptual framework with vague boundaries between MWM and PWM, disjointed reasoning mechanisms for the technical pathways and applicable scenarios of different Theory of Mind (ToM) reasoning paradigms, and detachment between evaluation and practice. To address these issues, this review systematically synthesizes over 100 authoritative studies to provide a comprehensive overview of MWM research for embodied AI. Its core contributions are threefold: First, it constructs a complete theoretical framework for MWM for the first time. Specifically, it distinguishes the essential differences between MWM and PWMs. Second, it systematically defines the key components of MWM through two paradigms for mental element representation. Third, it comprehensively analyzes two core ToM reasoning paradigms with 19 ToM methods. Finally, it also clarifies the integration trend of neuro-symbolic hybrid architectures, and synthesizes 26 ToM evaluation benchmarks. This work aims to promote the integration of embodied agents into human society and advance the in-depth development of human-machine collaborative interaction.

Figures (7)

• Figure 1: Framework of Mental World Model for Embodied AI Agents. (a) Concept clarification of world model, physical world model, and mental world model. (b) Organization of this review.
• Figure 2: Probabilistic Dependency Representation of State Variables in Physical and Mental World Models
• Figure 3: An external observation $o$ is first internally represented as $q(z|o)$ via a generative representation model, and then explicitly or implicitly decoded as $p(o'|z)$. The discrepancy between the decoded output $o'$ and the real observation $o$ is defined as the accuracy error. The second observation in the figure exaggeratedly illustrates the need to measure the complexity between the prior and the observed representation.$KL(q(z|o)|p(z))$.
• Figure 4: Embodied AI Agent Framework based on LLM and VLM. Left: theory of mind is inherently performed in human society, the LLM and VLM are trained based on human language and vision cues from evironment. Right: the embodied AI agent is contructed based on LLM/VLM, which forms the main capability of reasoning and planning.
• Figure 5: Typical Prompts of ToM Prompting Methods.