Reproduce “SceneGenAgent: Precise Industrial Scene Generation with Coding Agent” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Xiao Xia, Dan Zhang, Zibo Liao, Zhenyu Hou, Tianrui Sun, Jing Li, Ling Fu, Yuxiao Dong
SceneGenAgent provides a multi-stage, LLM-based coding agent for generating precise industrial 3D scenes from natural language descriptions. The system generates executable C# code for Siemens Tecnomatix, achieving up to 81.0% success with GPT-4o and improving fine-tuned open-source LLMs to 78.5% accuracy for quantitatively demanding industrial layouts.
Key insights
- A structured, multi-stage agentic approach with iterative verification is crucial for LLMs to achieve the high quantitative precision needed in industrial scene generation, overcoming limitations of single-call generation.
- Domain-specific instruction tuning, even with a moderately sized dataset like SceneInstruct, can significantly boost the performance of open-source LLMs for complex, precision-demanding tasks to levels comparable to advanced proprietary models.
- Explicit API coding guidance is essential for LLMs to generate correct and executable code for specialized industrial software, demonstrating the need for targeted framework-specific instructions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.