Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Toward Stable World Models: Measuring and Addressing World Instability in Generative Environments

Soonwoo Kwon, Jin-Young Kim, Hyojun Go, Kyungjune Baek

TL;DR

The paper defines world stability as the persistence of the initial scene after an agent executes an action sequence followed by its inverse, and introduces a WS score and evaluation framework to quantify this property in diffusion-based world models. It shows that state-of-the-art models struggle with stability and evaluates several improvement strategies—longer context, data augmentation, reverse prediction embedding, and refinement sampling—finding that reverse modeling and refinement yield notable gains while context length alone is not a universal solution. The proposed framework provides both a quantitative and qualitative lens to assess temporal-spatial consistency in generated worlds, with experiments on CS:GO and DMLab demonstrating the practical impact of stability on learning and user experience. Overall, world stability emerges as a vital criterion for reliable, safe, and scalable world modeling, guiding future research toward more temporally coherent generative environments and actionable strategies for improvement.

Abstract

We present a novel study on enhancing the capability of preserving the content in world models, focusing on a property we term World Stability. Recent diffusion-based generative models have advanced the synthesis of immersive and realistic environments that are pivotal for applications such as reinforcement learning and interactive game engines. However, while these models excel in quality and diversity, they often neglect the preservation of previously generated scenes over time--a shortfall that can introduce noise into agent learning and compromise performance in safety-critical settings. In this work, we introduce an evaluation framework that measures world stability by having world models perform a sequence of actions followed by their inverses to return to their initial viewpoint, thereby quantifying the consistency between the starting and ending observations. Our comprehensive assessment of state-of-the-art diffusion-based world models reveals significant challenges in achieving high world stability. Moreover, we investigate several improvement strategies to enhance world stability. Our results underscore the importance of world stability in world modeling and provide actionable insights for future research in this domain.

Toward Stable World Models: Measuring and Addressing World Instability in Generative Environments

TL;DR

The paper defines world stability as the persistence of the initial scene after an agent executes an action sequence followed by its inverse, and introduces a WS score and evaluation framework to quantify this property in diffusion-based world models. It shows that state-of-the-art models struggle with stability and evaluates several improvement strategies—longer context, data augmentation, reverse prediction embedding, and refinement sampling—finding that reverse modeling and refinement yield notable gains while context length alone is not a universal solution. The proposed framework provides both a quantitative and qualitative lens to assess temporal-spatial consistency in generated worlds, with experiments on CS:GO and DMLab demonstrating the practical impact of stability on learning and user experience. Overall, world stability emerges as a vital criterion for reliable, safe, and scalable world modeling, guiding future research toward more temporally coherent generative environments and actionable strategies for improvement.

Abstract

We present a novel study on enhancing the capability of preserving the content in world models, focusing on a property we term World Stability. Recent diffusion-based generative models have advanced the synthesis of immersive and realistic environments that are pivotal for applications such as reinforcement learning and interactive game engines. However, while these models excel in quality and diversity, they often neglect the preservation of previously generated scenes over time--a shortfall that can introduce noise into agent learning and compromise performance in safety-critical settings. In this work, we introduce an evaluation framework that measures world stability by having world models perform a sequence of actions followed by their inverses to return to their initial viewpoint, thereby quantifying the consistency between the starting and ending observations. Our comprehensive assessment of state-of-the-art diffusion-based world models reveals significant challenges in achieving high world stability. Moreover, we investigate several improvement strategies to enhance world stability. Our results underscore the importance of world stability in world modeling and provide actionable insights for future research in this domain.

Paper Structure

This paper contains 31 sections, 1 equation, 18 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Works
  3. World Simulation
  4. Evaluation Metrics for World Models
  5. Proposed Evaluation Framework
  6. Notation
  7. Evaluation Protocol
  8. World Stability Score
  9. Exploring Solutions for World Stability
  10. Longer Context Length
  11. Data Augmentation
  12. Inject Reverse Prediction Capability
  13. Refinement Sampling
  14. Experiments
  15. Experimental Setups
  16. ...and 16 more sections

Figures (18)

  • Figure 1: Revisiting the same place by a world model. When prompting the model to revisit the previously synthesized location, the newly generated environment fails to preserve key objects within the red box, such as a door or a picture frame. This inconsistency can be critically detrimental to both agent performance and user experience.
  • Figure 2: Baseline
  • Figure 3: Longer Context Length
  • Figure 4: Data Augmentation
  • Figure 5: Inject Reverse Prediction
  • ...and 13 more figures