Researchers have introduced WorldGuide, a goal-directed video world model that reframes procedural video generation as closed-loop task execution rather than a fixed open-loop rollout. The work appears as a paper on Hugging Face, with a project page and code released by the MBZUAI-ORYX group.
The core problem the paper identifies is that long-horizon procedural tasks require generation to adapt to what has actually been produced. A model must determine the next action from its generated state, execute that action, and recognize when the task is complete. Open-loop generation cannot adapt to execution outcomes, while existing closed-loop systems often depend on pretrained executors or indirect verification, leaving a gap between deciding an action and realizing it.
WorldGuide addresses this with two components trained on the same step-level procedural demonstrations: a ContextPlanner that predicts the next atomic action — or a completion token — from the generated visual state, and an Executor trained to render that action as a video clip. The generated result then guides the next action or termination. A hierarchical visual memory maintains state across long-horizon execution while keeping history token cost bounded.
Because step-level action-video supervision for joint planner-executor training was lacking, the team built WorldGuide Bench, comprising roughly 59K step-annotated videos spanning 245 tasks and 27 procedural categories.
On WorldGuide Bench, WorldGuide reaches 33.33% Task Success, compared with 29.90% for the recent video model MiniMax-H3 — even though MiniMax-H3 receives reference action plans. On VideoCraft-Bench, WorldGuide scores 47.69% versus 32.73% for MiniMax-H3 under goal-only conditioning. The authors conclude that coupling planning with learned execution is central to goal-directed procedural video generation.