Researchers at the University of Southern California introduced 3DCodeVerse, a research ecosystem designed to help coding agents generate editable, verifiable 3D content at scale. The project combines datasets of executable 3D programs, a coding framework for spatial reasoning and automated checks, and practical guidance for creating 3D content.

The team built -2M, a curated dataset of standalone 3D programs. It combines open-source material, converted specialized representations, and programs generated by coding agents. Each program includes the source files and assets needed to run it in its stated environment, plus descriptions and verification records when available.

The researchers also developed the Harness, a multi-agent directed acyclic graph framework that generates and revises 3D content. It checks program execution, measures spatial properties, and uses visual feedback to assess and revise programs without human feedback during generation. In experiments, fine-tuning open models on the dataset improved their 3D coding capabilities. The Harness achieved a 90% win rate against Claude Design (3D Object) in pairwise human evaluations when both were used by Claude Opus 5.5.

Researchers at the University of Southern California introduced 3DCodeVerse, a research ecosystem designed to help coding agents generate editable, verifiable 3D content at scale. The project combines datasets of executable 3D programs, a coding framework for spatial reasoning and automated checks, and practical guidance for creating 3D content.

The team built -2M, a curated dataset of standalone 3D programs. It combines open-source material, converted specialized representations, and programs generated by coding agents. Each program includes the source files and assets needed to run it in its stated environment, plus descriptions and verification records when available.

The researchers also developed the Harness, a multi-agent directed acyclic graph framework that generates and revises 3D content. It checks program execution, measures spatial properties, and uses visual feedback to assess and revise programs without human feedback during generation. In experiments, fine-tuning open models on the dataset improved their 3D coding capabilities. The Harness achieved a 90% win rate against Claude Design (3D Object) in pairwise human evaluations when both were used by Claude Opus 5.5.