NeurIPS 2025 Position Paper
Xiaoyuan Zhang1,2*, Chengdong Ma1*, Yizhe Huang1,2, Weidong Huang2,
Siyuan Qi2, Song-Chun Zhu1,2†, Xue Feng2†, Yaodong Yang1†
1PKU 2BIGAI
*Equal contribution †Equal corresponding authors
World models, which explicitly learn environmental dynamics to lay the foundation for planning, reasoning, and decision-making, are rapidly advancing in predicting both physical dynamics and aspects of social behavior, yet predominantly in separate silos. This division results in a systemic failure to model the crucial interplay between physical environments and social constructs, rendering current models fundamentally incapable of adequately addressing the true complexity of real-world systems where physical and social realities are inextricably intertwined. This position paper argues that the systematic, bidirectional unification of physical and social predictive capabilities is the next crucial frontier for world model development. We contend that comprehensive world models must holistically integrate objective physical laws with the subjective, evolving, and context-dependent nature of social dynamics. Such unification is paramount for AI to robustly navigate complex real-world challenges and achieve more generalizable intelligence. This paper substantiates this imperative by analyzing core impediments to integration, proposing foundational guiding principles (ACE Principles), and outlining a conceptual framework alongside a research roadmap towards truly holistic world models.
Physical & Social world model diagram summarizing
The social aspect is divided into Agent-Agent and Agent-Group Interaction, while the physical aspect distinguishes between interactive and non-interactive processes. Interactive dynamics are further classified into strategic and non-strategic interactions. Modalities are categorized into language, state, pixel, and 3D spaces.
The physical state evolution (governed by physical laws) and social state evolution (driven by Contingent Causality) are unified into a physical-social world model, which follows the ACE Principles, leading to impactful applications. Effective Abstraction is a prerequisite for understanding the Contingent Causality inherent in social dynamics. Both, in turn, are crucial for modeling the profound Entanglement between physical and social systems from which co-evolution and holistic emergence arise. Together, they form a cohesive set of guidelines for advancing towards truly comprehensive and predictive physical-social world models.
Regarding the technical barriers to scaling ACE: These principles directly map to core difficulties in modern AI. Abstraction targets the fundamental neural-symbolic grounding problem: bridging the gap between raw sensory data, where neural models excel, and the abstract, symbolic concepts like 'trust' that govern social dynamics. Contingent Causality addresses the problem of distributional shift in non-stationary social rules, a notorious weakness of current ML models. We propose that by incorporating inductive biases from behavioral economics regarding human decision-making and interaction patterns, models can better understand and predict how social rules are adopted and enacted in specific contexts. This fundamentally enhances the model's ability to discern 'which set of social rules applies here,' moving beyond mere statistical fitting.". Finally, Entangled Emergence focuses on the immense challenge of modeling the reciprocal causal influence between physical and social domains; modeling this continuous, bidirectional feedback loop remains a major hurdle.
Comparative overview of current world model benchmarks and their evaluation coverage, organized by physical-, social-, and unified-focus categories. The Brief Description column summarizes each benchmark’s main characteristics.
Three-tier hierarchical evaluation protocol for assessing socio-physical dynamics in world models. Each tier targets an increasingly integrated understanding of physical and social dynamics.
Comparison with Existing Positions and Surveys on World Models. Summarizes key focus areas across recent position papers and surveys.
This position paper has championed a pivotal paradigm shift for AI world models: the deep, bidirectional unification of their physical and social predictive capabilities. We argued that the prevalent separation of these dimensions renders current models fundamentally incomplete, hindering their capacity to capture the true complexity of our entangled physical-social reality and impeding progress towards AI systems that genuinely comprehend our multifaceted world.
To chart a constructive path forward, we delineated the distinct natures of physical and social predictions, underscored the imperative for their integration by highlighting their reciprocal interplay, critiqued existing formulations, and subsequently introduced three foundational ACE Principles. These principles, together with our proposed conceptual WM\textsubscript{P-S} framework and a research roadmap, collectively offer a structured, principled approach to developing world models that holistically represent and predict the co-evolution of physical and social realities.