WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
Video world models keep local quality but lose long-range memory while exploring. WorldProbe runs 1,600 rollouts across 10 models.
The claim
Video world models keep local quality but lose long-range memory while exploring.
The problem
Video world models synthesize worlds you can move through.
Existing benchmarks use short interactions or fixed camera paths.
Long-horizon exploration with heavy camera motion stays untested.
How it works
WorldProbe drives a goal-conditioned multimodal agent while frames are generated.
The agent adapts camera actions online as the world appears.
The benchmark holds 164 test cases grouped into four probing families.
Ten state-of-the-art models run more than 1,600 exploration rollouts.
The rollouts total about 800 minutes and reach horizons of 180 seconds.
Figures
What changes
Recent models sustain visual and content quality for longer stretches.
Long-range memory and evolution remain weak across the ten models.
Locally stable generation hides world inconsistencies that appear only at distance.
Limits
The study grades four dimensions over 164 fixed test cases.
Ten models and one agent policy frame every result reported here.
The results describe the current model releases, not future versions.
Source · arXiv