3DCodeVerse: 2M Executable 3D Code Dataset and Multi-Agent Coding Harness Achieve 90% Win Rate Against Claude Design

Background and the Problem

Generating 3D worlds as executable code offers unmatched editability, modularity, and structural interpretability compared to black-box mesh diffusion. However, coding agents struggle with spatial reasoning and execution reliability in 3D environments due to two critical bottlenecks: the scarcity of large-scale, self-contained executable 3D code repositories, and the absence of automated coding harnesses capable of verifying geometric dimensions, topological alignment, and rendered visual fidelity without human intervention.

Architecture and How It Works

Researchers from USC, University of Rochester, and NVIDIA introduced 3DCodeVerse, a comprehensive research ecosystem for programmatic 3D creation:

1. 3DCodeVerse-2M Executable Corpus: A curated dataset of 2 million standalone executable 3D programs built by aggregating open-source repositories, converting specialized 3D representations, and synthesizing novel programs via coding agents. Every entry is packaged with source code, required assets, execution environment specifications, and verification logs;
2. 3DCodeVerse Multi-Agent DAG Harness: A directed acyclic graph (DAG) framework that autonomously generates, evaluates, and revises 3D programs using three closed-loop verifiers: Execution Checks for runtime correctness, Spatial Measurements for quantitative geometric verification, and Visual Feedback for multi-view aesthetic and structural critique;
3. Open-Model Fine-Tuning & Component Guidance: Establishes rigorous baselines and component-level analyses showing how open-weight coding models can internalize 3D spatial programming.

Benchmarks and Measured Results

Empirical evaluations across open-source fine-tuning and frontier model comparisons highlight transformative gains:

• 90% Win Rate Against Claude Design: In pairwise human evaluations where both systems are powered by Claude Opus 5.5, the 3DCodeVerse Harness achieves a 90% win rate against Claude Design (3D Object);
• Dramatic Boost for Open-Weight Models: Fine-tuning open models on 3DCodeVerse-2M substantially improves executable compilation rates and spatial layout accuracy;
• Synergistic Spatial + Visual Verification: Ablation studies confirm that combining programmatic spatial measurements with rendered visual feedback eliminates floating artifacts and scale distortions.

Getting Started for Developers

The codebase and harness are open-sourced at https://github.com/gaoypeng/3dcodeverse:

• Automated 3D Asset & Scene Synthesis: Clone https://github.com/gaoypeng/3dcodeverse to run the multi-agent DAG harness for Three.js, Blender Python, and procedural 3D scene generation;
• Integrate Spatial Probes into Coding Harnesses: Pair VLM screenshot critiques with programmatic bounding-box and coordinate-intersection checks to catch 3D geometry bugs deterministically.