Reproduce “Frontier3DBench: Benchmarking Coding Agents for 3D Creation across Objects, Scenes, and Interactive Open Worlds” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Xinjie Lin, Qingle Liu, Wenyi Li, John Lam, Xiaoqiu Wang, Deyao Hong, Mingju Gao, Zheng Jiang and 5 more

Frontier3DBench evaluates coding agents on editable objects, scenes, and interactive worlds across 145 tasks. It combines rendered appearance, project inspection, and runtime tests. Configurations score lower on reference reconstruction than text-driven generation. Across 165 world submissions, local mechanisms average 83.08/100, versus 20.82/100 for motion clearance and 30.26/100 for stateful workflows. GPT-6.1-sol with Codex leads overall at 66.02/100.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.