LoopWM loops transformers for world models

LoopWM is interesting because it treats depth as reusable computation instead of more parameters, which is exactly the pressure point for learned simulators that must roll out thousands of imagined steps.

• –The core bet is that world dynamics benefit from iterative latent refinement, not just larger fixed-depth transformer stacks.
• –Spectral stability constraints are the right kind of boring engineering detail: without bounded recurrent updates, long-horizon rollout claims are easy to break.
• –Deferred decoding could matter for planning workloads because agents often need to score many candidate futures without rendering every intermediate state.
• –The 100x efficiency claim is still a paper result, not a production benchmark, so the next question is whether this holds outside curated text-heavy environments.
• –If replicated, this gives AI agent and robotics researchers a cleaner path to cheaper learned simulators than simply scaling video-world-model backbones.