Looped World Models

2026-06-16

AI Summary

Existing world models face a fundamental trade-off between computational depth and error accumulation in long-horizon simulation. This work addresses this limitation by introducing recurrent mechanisms into world modeling for the first time, proposing an approach that iteratively refines latent environment states through parameter-shared Transformer blocks. The number of iterative refinement steps—termed iterative latent state depth—is leveraged as a novel axis for model scaling. This design enables adaptive computation, dynamically adjusting the number of inference steps based on prediction complexity. Consequently, the method achieves high-fidelity simulation while attaining up to 100× greater parameter efficiency compared to conventional architectures.

Abstract

Current world models face a fundamental tension: faithful long-horizon simulation demands deep computation, but deeper models are expensive to deploy and prone to compounding errors. We resolve this by introducing Looped World Models (LoopWM), which are the first looped architectures for world modelling. Our method iteratively refines latent environment states through a parameter-shared transformer block. This yield up to 100x parameter efficiency over conventional approaches with adaptive computation that automatically scales depth to match the complexity of each prediction step. Orthogonal to scaling model size and training data, LoopWM establishes iterative latent depth as a new scaling axis for world simulation, which might significantly push the community forward.

Problem

• world models
• long-horizon simulation
• compounding errors
• parameter efficiency
• adaptive computation

Innovation

• Looped World Models
• iterative latent refinement
• parameter-shared transformer
• adaptive computation
• scaling axis