Looped World Models (LoopWM)
Summary
Looped World Models (LoopWM) is the world-model architecture introduced by Looped World Models. It uses a parameter-shared recurrent Transformer dynamics core to refine action-conditioned latent environment states with adaptive depth, spectral state-retention constraints, and an optional deferred-decoding rollout mode.
Role In The Wiki
Use this page as the object card for the LoopWM method family. The source page carries the evidence details, benchmark caveats, limitations, and agenda mapping.
Relation To Foundation TSFM Agenda
Use the source-level agenda mapping in looped-world-models-2026 rather than duplicating verdict rows here. At the entity level, LoopWM is relevant because it moves recurrent-depth Transformers from language-only latent reasoning into an action-conditioned world-model interface: latent state is rolled forward under actions, compute can vary by transition, and decoding can be deferred to a terminal state. The transfer target for TSFMs is adaptive latent-state update under a serving budget, not the specific text-game benchmark.
Evidence
Official Artifacts
- Preprint: arXiv 2606.18208
- Official X announcement: Hongyuan Adam Lu
- Review: ArXivIQ review
- Code status: no official repository was verified at ingest time; Gonzo/ArXivIQ list
Code: N/A. - Model status: no official checkpoint was verified at ingest time; Gonzo/ArXivIQ list
Model: N/A.