Alex Open Research Wiki

Fast-LeWM

2 min read

Fast-LeWM

Summary

Fast-LeWM is a compact action-conditioned JEPA world model that replaces LeWorldModel’s autoregressive one-step candidate rollout with causal action-prefix encoding and parallel multi-horizon latent prediction.

current observed latent + candidate control-input sequence
  -> causal prefix tokens
  -> parallel future latent states
  -> terminal goal cost + optional self-consistency
  -> CEM-selected action sequence

The main systems result is a reported 3.9× reduction in dynamics-module time, not a 4× end-to-end planner speedup. Full CEM time falls from 54.4 s to 28.3 s in the paper’s Two-Room measurement.

Interface

For each prefix length ,

Training supervises every horizon in the prefix block. Planning can use the terminal prefix directly instead of recursively generating all preceding imagined latent states.

Evidence Boundary

The source is arXiv v1 from 2026-06-24, not a verified peer-reviewed paper. The official repository contains a release-intent README but no public implementation at ingest time. Evidence covers four LeWM visual goal-reaching tasks with fixed , action skip 5, deterministic latent prediction, and CEM search. It does not yet establish variable-horizon scaling, stochastic futures, path-constraint handling, physical-robot transfer, or multivariate time-series control.

The user-provided X post is third-party commentary by Haiyu Wu. Its “4× CEM speedup” compresses a 3.9× dynamics-module speedup and a 48% full-CEM reduction; its Reacher and Cube gains use the optional self-consistency row.

Official Artifacts

Role In The Wiki

Fast-LeWM is the parallel multi-horizon dynamics-interface branch of the LeWorldModel line:

  • LeWorldModel supplies stable end-to-end JEPA state learning and one-step action-conditioned dynamics.
  • Fast-LeWM changes the action-sequence query interface and supervision so several horizons can be predicted in parallel.
  • Temporal Straightening changes planner-facing latent geometry instead; the two mechanisms are complementary hypotheses until compared jointly.
  • stable-worldmodel is the natural matched-protocol reproduction surface once implementation code is released.

Relation To Foundation TSFM Agenda

Use the source-level agenda mapping in Fast LeWorldModel rather than duplicating slot verdicts here.

For time-series/world-model research, the transferable object is a causal encoder over candidate control-input or intervention prefixes plus a parallel multi-horizon state predictor. A serious transfer must add persistent state for partial observability, irregular-time handling, typed action status, path constraints, uncertainty, and event/regime preservation.

Graph View

Backlinks