SmallWorlds: Assessing Dynamics Understanding of World Models in Isolated Environments

TL;DR

This paper introduces the SmallWorld Benchmark, a controlled test environment for evaluating world models' ability to learn environment dynamics, providing insights into their strengths and limitations across various architectures.

Contribution

The paper presents the SmallWorld Benchmark for systematic evaluation of world models, enabling controlled assessment of their dynamic understanding without reward signals.

Findings

Models capture environment structure to varying degrees

Prediction quality deteriorates over long rollouts

Different architectures exhibit distinct strengths and weaknesses

Abstract

Current world models lack a unified and controlled setting for systematic evaluation, making it difficult to assess whether they truly capture the underlying rules that govern environment dynamics. In this work, we address this open challenge by introducing the SmallWorld Benchmark, a testbed designed to assess world model capability under isolated and precisely controlled dynamics without relying on handcrafted reward signals. Using this benchmark, we conduct comprehensive experiments in the fully observable state space on representative architectures including Recurrent State Space Model, Transformer, Diffusion model, and Neural ODE, examining their behavior across six distinct domains. The experimental results reveal how effectively these models capture environment structure and how their predictions deteriorate over extended rollouts, highlighting both the strengths and limitations of current modeling paradigms and offering insights into future improvement directions in representation learning and dynamics modeling.