Looped World Models

The paper introduces Looped World Models (LoopWM), the first looped transformer architecture for world modeling, addressing the tension between deep computation for faithful long-horizon simulation and the high cost and error accumulation of deep models.

LoopWM iteratively refines latent environment states through a parameter-shared transformer block with spectrally-constrained residual dynamics, achieving up to 100× parameter efficiency over conventional approaches.

The architecture includes a prelude, a recurrent block with shared parameters, and a coda, with stability guaranteed by constraining the state-retention matrix's eigenvalues to (0,1).

Training uses stochastic loop depth and entropy-regularized adaptive early exit, enabling test-time compute scaling.

A deferred decoding variant eliminates intermediate observation decoding, producing predictions only at the terminal step, with a curriculum over the deferral horizon.

On ScienceWorld, LoopWM (about 1B parameters) outperforms claude-opus-4-6-max by 21.2% in EM (68.4% vs 47.2%) and shows strong results on AlfWorld, despite being over 100× smaller.

The work establishes iterative latent depth as a new scaling axis for world simulation, orthogonal to model size and data volume, with potential for significant community impact.

The main limitation is presentation scope, with broader validation and scaling analysis deferred to future work.

The paper also includes human evaluation results on danmaku generation, where LoopWM outperforms baselines on appropriateness, informativeness, engagement, and human-likeness.

Overall, LoopWM demonstrates that looped architectures can provide stable, efficient, and adaptive world models, opening new directions for building more capable and efficient world simulators.

The authors believe this work identifies iterative latent depth as a new scaling axis for world…