3DCodeBench: Benchmarking Agentic Procedural 3D Modeling Via Code

Researchers introduce 3DCodeBench, a comprehensive benchmark for evaluating vision-language models (VLMs) as procedural 3D modelers that convert text and image inputs into code for 3D modeling software. The study reveals that current advanced VLMs struggle primarily with API mismatches and geometric coherence, while identifying test-time scaling as an effective improvement method.

3DCodeBench addresses a critical gap in AI evaluation by systematically testing how well vision-language models can perform procedural 3D modeling—a task requiring both creative interpretation and precise technical execution. The benchmark is significant because procedural code-based 3D modeling offers advantages over neural generation methods: assets are deterministic, engine-ready, and easily editable. This positions procedural approaches as commercially viable for game development, CAD systems, and digital asset creation.

The research reveals fundamental limitations in current VLMs. Rather than fundamental reasoning gaps, failures stem from practical issues: incorrect API calls, parameter mismatches, and geometric errors like disconnected or floating components. This suggests the problem is not insurmountable—better training data and API-specific fine-tuning could substantially improve performance. The introduction of 3DCodeArena, a human-preference-based ranking platform, acknowledges that automated metrics inadequately capture perceptual 3D quality, reflecting broader challenges in AI evaluation methodology.

For the AI development community, 3DCodeBench establishes evaluation standards that will drive improvement in VLM capabilities for specialized domains. The finding that test-time scaling (extended thinking budgets and multi-turn refinement) improves results suggests computational investment can partially offset training limitations. The release of curated datasets, evaluation protocols, and public evaluation infrastructure creates shared benchmarks that accelerate progress.

Looking forward, this research indicates the next frontier for VLM advancement lies in domain-specific training and robust execution feedback loops. Commercial AI providers investing in procedural 3D modeling will require specialized datasets and tighter integration with 3D software ecosystems to achieve production-ready performance.

• →3DCodeBench provides the first systematic evaluation framework for vision-language models in procedural 3D modeling tasks.
• →Current VLM failures predominantly stem from API mismatches rather than fundamental reasoning limitations, suggesting targeted improvements are achievable.
• →Test-time scaling and iterative refinement substantially improve VLM performance on procedural 3D generation tasks.
• →The research highlights critical demand for high-quality procedural coding datasets to advance commercial VLM capabilities.
• →Human preference-based evaluation (3DCodeArena) proves more reliable than automated metrics for assessing 3D output quality.