3DCodeVerse pairs 2M 3D programs with verification harness

3DCodeVerse combines a 2M executable-3D dataset and a verification harness that wins 90% against Claude Design.

3DCodeVerse is a research ecosystem for generating editable, verifiable 3D content as executable code. It pairs 3DCodeVerse-2M, a curated corpus of standalone executable 3D programs with assets and verification records, with a multi-agent DAG harness that revises outputs using execution checks, spatial measurements, and visual feedback. The corpus draws on open-source data, converted representations, and agent-generated programs, and one report states that harness revisions do not rely on human feedback. Fine-tuning on the dataset substantially improves open models’ 3D coding, and the authors report baselines plus an analysis of harness components. In pairwise human evaluations, the harness used with Claude Opus 5.5 achieved a 90% win rate against Claude Design for 3D objects; the earlier report specifies that both sides used Claude Opus 5.5, while the later report describes the model as used with the harness.

• 3DCodeVerse is a research ecosystem for coding agents that generate editable, verifiable 3D content as executable code.
• 3DCodeVerse-2M is a curated corpus of about 2 million standalone executable 3D programs, packaged with assets and verification records, built from open-source data, converted representations, and agent-generated code.
• A multi-agent DAG harness revises programs using execution checks, spatial measurements, and visual feedback, without human feedback during revision.
• Fine-tuning on the dataset substantially improves open models’ 3D coding; the authors also set baselines with fine-tuned open models and analyze harness components.
• With Claude Opus 5.5, the harness reached a 90% pairwise human-evaluation win rate against Claude Design for 3D objects; one report says both sides used Claude Opus 5.5.
• Coverage dates are Hugging Face daily papers on 2026-10-14 and arXiv (cs.AI / cs.LG / cs.CL) on 2026-10-15.