Code4Scene benchmark: coding agents still fail at building and editing Unreal Engine 3D scenes

Lianhuiq · x · 2026-10-01

Researchers introduce Code4Scene, a benchmark testing coding agents on constructing and editing Unreal Engine 3D scenes via code, evaluated across 14 agent configurations on 95 cases. Findings: spatial composition is the weakest construction skill for every agent, and 35.8% of edits that fully recover the target still introduce unintended changes elsewhere.

Original post →

More from coding & agent

coding & agent channel →