SimWorld Benchmark: 35.8% of 3D Scene Edits Break Something Else

ZhitingHu · x · 2026-10-01

The SimWorld team releases SimWorld-Code4Scene: 190 Unreal Engine cases testing spatial reasoning via text-to-scene construction and precise control via image-guided editing. Unlike screenshot-based evaluation, they reopen the actual generated 3D scenes and assess instruction following, physical validity, and scene integrity. Astra leads overall, but spatial composition is the weakest construction category across all 14 tested configurations, and 35.8% of edits that fully recover their targets still introduce unintended changes elsewhere. Paper and code are available.

Related event: Code4Scene Benchmark Tests Agents Building 3D Scenes in Unreal Engine(2 posts)→

Original post →

More from Multimodal

Multimodal channel →