Looped World Models match billion-parameter systems with 100× fewer parameters

Looped World Models (LoopWM), introduced in a preprint posted to arXiv on June 20, is a recurrent-depth transformer architecture that models environment dynamics by iteratively refining latent representations through a single shared-weight block. The work introduces a mathematically guaranteed contractive state-retention mechanism and an adaptive early-exit strategy that lets the model trade inference speed for accuracy without retraining. A 1-billion-parameter LoopWM checkpoint outperforms larger proprietary world models on long-horizon planning benchmarks, with the authors claiming parameter efficiency gains up to 100×.

Classical world models face a rigid tradeoff: modeling long planning horizons demands deep, parameter-heavy architectures that accumulate rollout errors and strain deployment budgets. LoopWM sidesteps that constraint by treating iterative latent depth as an orthogonal scaling axis. Instead of stacking dozens of unique transformer layers, the architecture loops a single block with shared weights, refining the environment's latent state on each pass. The contractive state-retention mechanism ensures convergence, and early-exit logic halts the loop when additional iterations yield diminishing returns. That design gives practitioners a dial to tune compute cost at inference time without retraining.

The architecture extends to Deferred Decoding (LoopWM-DD), which rolls out entire action trajectories in latent space and decodes only the final step. Autoregressive world models decode every intermediate frame, compounding errors at each step and burning cycles on pixels that planning algorithms never inspect. By deferring decoding to the end of a rollout, LoopWM-DD speeds up physics simulation for reinforcement learning and fits tighter power and latency budgets on edge robotics platforms. Code and model weights have not been released; the full preprint is available at arXiv:2606.18208.