WorldGuide
A closed-loop video world model framework in which a planner and an executor are jointly trained to carry out procedural long-horizon tasks in visual space. It supports atomic action prediction, video segment generation and next-step decision-making based on feedback from generated results.
Timeline (last 90 days)
WorldGuide achieved a 33.33% task success rate on WorldGuide-Bench.
WorldGuide reached a 47.69% success rate on VideoCraft-Bench.
WorldGuide's architecture consists of two core components, a Planner and an Executor, which are jointly trained on the same step-level procedural demonstration data.
WorldGuide uses hierarchical visual memory to maintain state during long-horizon execution, controlling token cost while preserving historical information.
WorldGuide defines procedural video generation as closed-loop task execution in a visual world space: given an initial image and a task goal, the model predicts atomic actions, generates video clips, and uses the generated results to choose the next action or determine task completion.
Existing video procedural task systems either adopt open-loop generation that cannot adjust actions based on execution results, or rely on pretrained executors or indirect verification, creating a gap between the action decided and the action realized.
WorldGuide's experimental results show that coupling planning with learnable execution is key to goal-directed procedural video generation.
WorldGuide demonstrates the feasibility of a closed-loop plan-execute architecture in procedural video tasks, but the 33.33% task success rate shows that issues such as long-horizon visual memory efficiency, atomic action granularity matching, and cross-category generalization ability remain unresolved.
MiniMax-H3 received an additional action plan input in the comparison while WorldGuide did not, indicating that merely providing action planning information is not enough to guarantee task completion, and whether the executor can faithfully realize the plan is where the bottleneck lies.
All Facts (9)
WorldGuide achieved a 33.33% task success rate on WorldGuide-Bench.
50%UnverifiedWorldGuide reached a 47.69% success rate on VideoCraft-Bench.
50%UnverifiedWorldGuide's architecture consists of two core components, a Planner and an Executor, which are jointly trained on the same step-level procedural demonstration data.
50%UnverifiedWorldGuide uses hierarchical visual memory to maintain state during long-horizon execution, controlling token cost while preserving historical information.
50%UnverifiedWorldGuide defines procedural video generation as closed-loop task execution in a visual world space: given an initial image and a task goal, the model predicts atomic actions, generates video clips, and uses the generated results to choose the next action or determine task completion.
50%UnverifiedExisting video procedural task systems either adopt open-loop generation that cannot adjust actions based on execution results, or rely on pretrained executors or indirect verification, creating a gap between the action decided and the action realized.
50%UnverifiedWorldGuide's experimental results show that coupling planning with learnable execution is key to goal-directed procedural video generation.
50%UnverifiedWorldGuide demonstrates the feasibility of a closed-loop plan-execute architecture in procedural video tasks, but the 33.33% task success rate shows that issues such as long-horizon visual memory efficiency, atomic action granularity matching, and cross-category generalization ability remain unresolved.
50%UnverifiedMiniMax-H3 received an additional action plan input in the comparison while WorldGuide did not, indicating that merely providing action planning information is not enough to guarantee task completion, and whether the executor can faithfully realize the plan is where the bottleneck lies.
50%