3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code

Researchers introduced 3DCodeBench, a benchmark for testing 12 advanced vision-language models on text- and image-to-procedural-3D code generation.

They also launched 3DCodeArena, a human-preference ranking platform to judge the quality of generated 3D outputs.

Evaluations found common API mismatch errors and geometric flaws even when renders succeeded, showing the task remains difficult.

More test-time reasoning and iterative refinement improved results, and the team released the dataset, evaluation protocol, and public platform.