3DCodeVerse: Exploring 3D Worlds as Code

3DCodeVerse. Exploring 3D Worlds as Code. By Yipeng Gao and eighteen others, from the University of Southern California, Yale University, and seven other institutions. Published on the fifteenth of October. Its code repository is worked out: about eight hundred and fifty source files, with a readme and a licence.

3DCodeVerse is a research ecosystem built so that coding agents can create 3D content that is editable and verifiable, at scale. Its premise is in the subtitle: treat 3D worlds as code. The ecosystem has three parts. Large, high-quality datasets of executable 3D code. A coding harness for spatial reasoning and automated verification. And practical guidance for 3D coding, drawn from systematic studies.

The centrepiece dataset is 3DCodeVerse-2M, a curated collection of standalone, executable 3D code. Each program comes packaged with what it needs to run: the source files and associated assets required for execution in its declared environment. Alongside those travel whatever descriptions are available, and verification records. A program in the collection is meant to be something that can actually be run where it says it runs, with a record of whether it was checked, rather than a loose snippet.

The second part is the 3DCodeVerse Harness, a multi-agent framework organised as a directed acyclic graph, for generating and refining 3D content. It evaluates and revises generated programs automatically, using execution checks, spatial measurements and visual feedback, with no human feedback needed while generation is under way. Does the program run, do the spatial measurements of the scene hold up, and does the result look right: those three questions drive each revision.

The corpus was assembled three ways: by collecting open-source data, by converting specialised representations into code, and by generating new programs with coding agents. The guidance comes from systematic empirical studies of 3D coding, which establish baselines with fine-tuned open models and analyse the harness one component at a time. So the dataset is part gathered, part translated, and part written by the same kind of agent the ecosystem is meant to serve.

For anyone training models to write 3D code, the reported result is that fine-tuning on the dataset significantly improves the 3D coding capabilities of open-source models. For anyone building agents, the harness carries the headline number. In pairwise human evaluations it achieves a ninety percent win rate against Claude Design, 3D Object, when both are driven by Claude Opus 5.5.

The ecosystem rests on one property of code. A program can be executed, measured and looked at, and those are exactly the three checks the harness runs on its own output. Data to learn from, a loop to verify with, and guidance on how to use both arrive as one package.