Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Generative Engines to Actionable Simulators: The Imperative of Physical Grounding in World Models

Zhikang Chen, Tingting Zhu

TL;DR

The paper tackles the gap between high-fidelity visual generation and true world understanding by reframing world models as actionable simulators rather than purely perceptual engines. It advocates structured $4D$ interfaces, physics-informed grounding, and closed-loop evaluation to sustain causal, long-horizon reasoning, with medicine highlighted as a critical stress test for safety and decision-making. The authors synthesize approaches across external interfaces, self-evolution, and domain-grounded evaluation, introducing benchmarks like WorldModelBench and WorldEval while showcasing MeWM and RoboNurse-VLA as domain-specific demonstrations. The study argues that robust, counterfactual planning and intervention capabilities are the practical yardsticks for true world models, outlining a path toward the integration of causality, physics, and adaptive learning as the foundation of the next generation of intelligent systems.

Abstract

A world model is an AI system that simulates how an environment evolves under actions, enabling planning through imagined futures rather than reactive perception. Current world models, however, suffer from visual conflation: the mistaken assumption that high-fidelity video generation implies an understanding of physical and causal dynamics. We show that while modern models excel at predicting pixels, they frequently violate invariant constraints, fail under intervention, and break down in safety-critical decision-making. This survey argues that visual realism is an unreliable proxy for world understanding. Instead, effective world models must encode causal structure, respect domain-specific constraints, and remain stable over long horizons. We propose a reframing of world models as actionable simulators rather than visual engines, emphasizing structured 4D interfaces, constraint-aware dynamics, and closed-loop evaluation. Using medical decision-making as an epistemic stress test, where trial-and-error is impossible and errors are irreversible, we demonstrate that a world model's value is determined not by how realistic its rollouts appear, but by its ability to support counterfactual reasoning, intervention planning, and robust long-horizon foresight.

From Generative Engines to Actionable Simulators: The Imperative of Physical Grounding in World Models

TL;DR

The paper tackles the gap between high-fidelity visual generation and true world understanding by reframing world models as actionable simulators rather than purely perceptual engines. It advocates structured interfaces, physics-informed grounding, and closed-loop evaluation to sustain causal, long-horizon reasoning, with medicine highlighted as a critical stress test for safety and decision-making. The authors synthesize approaches across external interfaces, self-evolution, and domain-grounded evaluation, introducing benchmarks like WorldModelBench and WorldEval while showcasing MeWM and RoboNurse-VLA as domain-specific demonstrations. The study argues that robust, counterfactual planning and intervention capabilities are the practical yardsticks for true world models, outlining a path toward the integration of causality, physics, and adaptive learning as the foundation of the next generation of intelligent systems.

Abstract

A world model is an AI system that simulates how an environment evolves under actions, enabling planning through imagined futures rather than reactive perception. Current world models, however, suffer from visual conflation: the mistaken assumption that high-fidelity video generation implies an understanding of physical and causal dynamics. We show that while modern models excel at predicting pixels, they frequently violate invariant constraints, fail under intervention, and break down in safety-critical decision-making. This survey argues that visual realism is an unreliable proxy for world understanding. Instead, effective world models must encode causal structure, respect domain-specific constraints, and remain stable over long horizons. We propose a reframing of world models as actionable simulators rather than visual engines, emphasizing structured 4D interfaces, constraint-aware dynamics, and closed-loop evaluation. Using medical decision-making as an epistemic stress test, where trial-and-error is impossible and errors are irreversible, we demonstrate that a world model's value is determined not by how realistic its rollouts appear, but by its ability to support counterfactual reasoning, intervention planning, and robust long-horizon foresight.
Paper Structure (8 sections, 1 figure, 6 tables)

This paper contains 8 sections, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. External Interfaces of World Models: From Generation to Long-Horizon and Causal Futures
  3. Self-Evolution and Long-Term World Consistency
  4. Self-Evolution with Physical Anchors: Physics-Informed World Models
  5. Generalization under Limited Interaction: Learning in Imagination
  6. World Models in Medical Settings: The Ultimate Epistemic Stress Test
  7. Evaluation: When Does a Model Qualify as a World Model?
  8. Conclusion: The Imagination Engine of General Intelligence

Figures (1)

  • Figure 1: Direct interaction between robots (VLA) and the real world is inherently limited due to safety, cost, and efficiency constraints. A world model enables imagination-based learning by simulating environment dynamics, physical laws, and state transitions in a latent space. Through self-supervised generation, generalization, and evaluation within the world model, the agent can acquire richer knowledge beyond observed data and effectively transfer the learned knowledge and policies to real-world tasks.