3DCodeVerse

3DCodeVerse is a combined data, execution and evaluation ecosystem that treats 3D worlds as executable code. It packages 269,597 runnable programs expanded into 2.03 million code-condition pairs spanning seven languages/runtimes (bpy, CadQuery, OpenSCAD, Three.js, URDF, GLSL, OpenGL) and four task tracks (objects, articulated objects, scenes, graphics). Each program ships with sources, assets, multi-view renders, three caption types (instruction, visual description, construction procedure) and a verification record. A web showcase provides 24 code-authored worlds and live in-browser examples; generated assets are demonstrated in downstream uses such as robot simulation and video generation.

The harness is an LLM-planned, DAG-orchestrated pipeline that builds, executes, measures and judges programs, then applies targeted automated repairs using spatial measurements (contacts, collisions, distances) and a vision-language judge. Iterating adversarially on generator and judge yields large quality and run-rate gains: the same model inside this harness wins most head-to-head comparisons, with human-preference increases often approaching ~90% over native toolchains. Fine-tuning on 3DCodeVerse-2M raised Qwen3.8-27B execution on a 212-task benchmark from 51.9% to 92.9%; per-runtime execution after tuning reached high rates (e.g., CadQuery ~98.5%, Blender ~93%). Datasets, harness code and models are released for research.