3DCodeVerse pairs a large 3D-code dataset with a multi-agent harness
An arXiv preprint introduces 3DCodeVerse, a research ecosystem built around 3DCodeVerse-2M, a large-scale curated dataset of standalone executable 3D code, plus a 3DCodeVerse Harness for spatial reasoning and automated verification. The harness is a multi-agent directed acyclic graph framework that uses execution checks, spatial measurements and visual feedback to evaluate and revise generated programs without human feedback during generation, the authors say. They also report that fine-tuning on the dataset improves open-source models' 3D coding and that the harness won 90% of pairwise human evaluations against Claude Design (3D Object) when both were used by Claude Opus 5.5; these results are self-reported and not independently replicated.
The authors present the dataset and harness as infrastructure for coding agents to create editable, verifiable 3D content at scale.
Computer Science > Computer Vision and Pattern Recognition
arXiv:2610.22341 (cs)
[Submitted on 15 Oct 2026]
Title:3DCodeVerse: Exploring 3D Worlds as Code
Authors:Yipeng Gao, Ziyao Zeng, Jinfa Huang, Yijiang Li, Haolin Xiong, Wang Qin, Xinlei Yu, Youheng Yao, Jie Zhang, Kun-Yu Lin, Rong Liu, Yinyuan Zhao, Yajie Zhao, Meiqi Guo, Jiebo Luo, Lei Shu, Yunhao Ge, Xinyu Hu, Laurent Itti
View a PDF of the paper titled 3DCodeVerse: Exploring 3D Worlds as Code, by Yipeng Gao and 18 other authors
#3DCodeVerse#3DCodeVerse-2M#3DCodeVerse Harness#Claude Opus 5.5#Claude Design