WorldGuide: A Goal-Oriented Video World Model for Executing Procedural Tasks

MBZUAI-ORYX has unveiled 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 system pairs a ContextPlanner, which predicts the next atomic action or a completion token from the generated visual state, with an Executor trained on the same step-level demonstrations to render that action as a clip. The rendered result then guides the next action or triggers termination, with hierarchical visual memory keeping long rollouts at bounded token cost.

Background: Why Step-Level Supervision Was Missing

Most video world models generate a fixed sequence of frames without checking whether the task is actually progressing. Procedural tasks — assembling, cooking, operating equipment — require the model to decide what to do next based on what it has already produced. WorldGuide addresses this by treating generation as a loop: plan an atomic action, render it, observe the new state, repeat or stop. To train such a loop, the team needed step-level action-video supervision that existing datasets did not provide.

Development: WorldGuide Bench and the Executor

To fill that gap, the researchers built WorldGuide Bench, comprising roughly 59K step-annotated videos spanning 245 tasks across 27 procedural categories. The Executor is trained on these same step-level demonstrations, so the clip it renders corresponds to a specific predicted action rather than a generic continuation.

Latest Results: Task Success Rates

On WorldGuide Bench, WorldGuide reaches 33.33% Task Success, compared with 29.90% for MiniMax-H3, which receives reference action plans. On Video-CraftBench, WorldGuide reaches 47.69%, versus 32.73% under goal-only conditioning. The project page is available at https://mbzuai-oryx.github.io/WorldGuide, and the code is at https://github.com/mbzuai-oryx/WorldGuide.

What to Watch Next

The comparison against MiniMax-H3 is notable because that baseline is given reference action plans, while WorldGuide must predict its own next action from generated visual state. The gap on Video-CraftBench — 47.69% versus 32.73% — suggests that closed-loop planning with step-level supervision outperforms goal-only conditioning on procedural tasks. Whether hierarchical visual memory scales to longer horizons and more categories than the 27 in WorldGuide Bench remains an open question for follow-up work.