An MBZUAI team built WorldGuide, a system that uses each generated video clip to choose the next action in a multistep task. On its WorldGuide Bench, the full system reached 33.33% task success, compared with 11.71% for a version that fixed all actions before generating. Gemini 3.5 Flash judged whether each final video visibly completed the task; this was not a physical robot test.
Figure 1 from Ankan Deria et al., “WorldGuide”, licensed under CC BY 4.0.
AI News | Research
Given an initial image and a task goal, WorldGuide’s planner proposes one action at a time. A video generator renders a short clip, and the system uses that clip to choose the next action or stop. The planner builds on Qwen2.5-VL-7B-Instruct; HunyuanVideo-1.5 generates the clips.
The researchers built WorldGuide Bench from 58,679 videos annotated with individual steps, spanning 245 tasks and 27 procedural categories. The paper lists public release of the dataset as planned.
On WorldGuide Bench, WorldGuide scored 33.33% with its own actions, while MiniMax-H3 scored 29.90% after receiving reference action plans. The project page says the 3.43-percentage-point difference is not statistically significant (p=0.10) and notes the input conditions differ. On Video-CraftBench, both models received only a starting image and task goal; the paper reports video-judged task-success scores of 47.69% for WorldGuide and 32.73% for MiniMax-H3.
The system’s feedback loop still inherits errors from its own generated videos. The authors report failures when the planner acts too early or repeats an action, or when a clip does not fully realize the requested step; any of those can mislead later decisions.