WorldGuide: Goal-Directed Video World Model for Procedural Task Execution

TL;DR

WorldGuide is a closed-loop video generation system that plans and executes procedural tasks using visual feedback from generated video clips. It combines a planner and executor trained on step-level demonstrations, using hierarchical visual memory to track state efficiently. The model achieves higher task success than prior methods without relying on external action plans.

Why It Matters

This approach enables more reliable automation of complex, long-horizon visual tasks by closing the loop between planning and execution, reducing reliance on pre-specified action sequences. It advances the feasibility of AI agents that can learn and adapt through visual feedback in real-world-like environments.

Key Findings

• WorldGuide outperforms MiniMax-H3 by 3.43 percentage points on WorldGuide-Bench and 14.96 points on VideoCraft-Bench under goal-only conditioning.
• The system uses a joint planner-executor architecture trained on step-level demonstrations, enabling direct action realization from visual progress.
• Hierarchical visual memory maintains state across long sequences with low computational overhead, supporting scalable procedural execution.