VLWM

Summary

VLWM is the Vision Language World Model introduced by Planning with Reasoning using Vision Language World Model. It predicts high-level goals and interleaved textual action/state-change trajectories from video context, then supports either direct System-1 decoding or System-2 candidate-plan selection with a learned semantic critic.

Model Interface

visual context
  -> goal + initial/final-state interpretation
  -> action / textual state-change trajectory
  -> direct plan or critic-ranked candidate plan

VLWM should be treated as a high-level semantic world model and planner, not as a low-level robot controller. Its actions are natural-language procedural steps rather than continuous control inputs, and its world state is free-form text rather than a calibrated physical-state representation.

Official Artifacts

As checked on 2026-07-15, the model page is manually gated and exposes no public weight file in its public manifest; the dataset repository has zero payload bytes; and no official code repository could be verified. The ICLR 2026 submission is marked withdrawn.

Relationship To Action100M

Action100M is a later hierarchical action/caption release by an overlapping team whose paper says it extends and improves the VLWM data-generation pipeline. It is not the same dataset or target schema as VLWM’s unreleased goal/action/state-change corpus.

Role In The Wiki

VLWM is the main language-state branch of the world-model map. It provides a concrete alternative to both pixel rollout and opaque latent rollout: predict readable semantic state changes and use a semantic cost model for high-level planning.

Its main warning is equally important. Readability does not guarantee physical sufficiency. A textual state can omit geometry, timing, stochasticity, partial observability, constraints, and continuous control details.

Relation To Foundation TSFM Agenda

Use the source-level slot mapping in VLWM. At entity level, VLWM is an adjacent architecture pattern for separating high-level action/state reasoning from low-level signal and control layers; it is not numeric time-series evidence.