Four Principles for Physically Interpretable World Models

2025-03-04

AI Summary

In open, dynamic environments, autonomous systems require trustworthy world models; however, existing approaches lack explicit correspondence between latent variables and physical quantities/dynamics, hindering interpretability, formal safety verification, and downstream control. Method: We propose the “physically interpretable world model” paradigm, formally defining physical interpretability beyond mere physics-informed learning, and establishing four principles: (i) symbolic-knowledge-guided latent space structuring, (ii) symmetry-aligned representation learning, (iii) multi-granularity supervised adaptation, and (iv) modularized generation. Our approach integrates invariant/equivariant representation learning, hierarchical supervision, and a modular generative architecture. Contribution/Results: Empirical evaluation on two benchmarks demonstrates that each principle significantly improves interpretability, cross-scenario generalization, and efficiency of formal verification—providing both a theoretical framework and a practical technical pathway toward fully physically interpretable world models.

Technology Category

• Computer Vision: Low Level & Physics-based Vision
• Reasoning under Uncertainty: Relational Probabilistic Models
• Cognitive Modeling & Cognitive Systems: Symbolic Representations

Application Category

• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

As autonomous systems are increasingly deployed in open and uncertain settings, there is a growing need for trustworthy world models that can reliably predict future high-dimensional observations. The learned latent representations in world models lack direct mapping to meaningful physical quantities and dynamics, limiting their utility and interpretability in downstream planning, control, and safety verification. In this paper, we argue for a fundamental shift from physically informed to physically interpretable world models - and crystallize four principles that leverage symbolic knowledge to achieve these ends: (1) structuring latent spaces according to the physical intent of variables, (2) learning aligned invariant and equivariant representations of the physical world, (3) adapting training to the varied granularity of supervision signals, and (4) partitioning generative outputs to support scalability and verifiability. We experimentally demonstrate the value of each principle on two benchmarks. This paper opens several intriguing research directions to achieve and capitalize on full physical interpretability in world models.

Problem

• Need for trustworthy world models in uncertain environments
• Lack of interpretable latent representations in current models
• Proposing principles for physically interpretable world models

Innovation

• Structured latent spaces by physical intent
• Aligned invariant and equivariant representations
• Adapted training for varied supervision granularity