A new open-world benchmark called WorldRoamBench finds that no current interactive world model holds up reliably under extended, continuous interaction.
Researchers introduced WorldRoamBench to address a blind spot in how interactive world models are evaluated. Prior benchmarks measured action-following only at the trajectory level, missing failures that emerge mid-sequence or across longer time horizons. WorldRoamBench tests across four dimensions: action accuracy on a per-frame basis, visual drift that catches non-monotonic collapse mid-sequence, physics plausibility covering mechanics and 3D consistency, and scene and subject memory tested through point-cloud reconstruction and vision-language reasoning. The benchmark runs more than 600 test cases across nature, urban, and indoor environments in both first- and third-person views, with 10 to 60 seconds of continuous WASD interaction.
The results matter because interactive world models are increasingly pitched as the substrate for embodied AI, game simulation, and robotics - applications where a model that forgets a room layout or invents physics mid-run is not a curiosity but a failure. Evaluating more than 10 open- and closed-source models, the researchers found that even the best performer hit only moderate scores, and none satisfied all four dimensions reliably.
That "none reliably satisfies all dimensions" finding is the kind of result that should temper the hype around world models as deployment-ready infrastructure - and suggests the gap between impressive demos and stable, memory-faithful systems is wider than the lab announcements tend to admit.