Director, ArchiAI Lab · Arcplus Group
Ph.D. in Civil Engineering, Tongji University — working on AI for Engineering
I lead the ArchiAI Lab at Arcplus Group in Shanghai, a cross-disciplinary team building AI for the built environment. My research asks one question: can machines read engineering drawings the way engineers do?
Toward that goal, I work on vector multimodal models for CAD drawings, large-scale open datasets (ArchCAD-400K, NeurIPS 2025), and domain benchmarks for large language models (AECBench, Advanced Engineering Informatics). Earlier, my Ph.D. work at Tongji University applied reinforcement learning and Monte-Carlo tree search to structural design. I also serve as an adjunct graduate supervisor at Tongji University.
Teaching models to parse CAD natively — panoptic symbol spotting, vector-line representations, and the data engines that make large-scale annotation affordable.
Hierarchical benchmarks that measure what engineering practice actually demands: knowledge memorization, understanding, reasoning, calculation, and application.
Reinforcement learning and Monte-Carlo tree search for generative structural design, from truss layout to topology optimization.
AECBench published in Advanced Engineering Informatics (Vol. 71).
Two papers accepted at NeurIPS 2025: ArchCAD-400K and “Point or Line?”.
ArchCAD-400K dataset released on GitHub and Hugging Face — 413K annotations, 5,538 drawings.
413K annotations across 27 categories from 5,538 standardized drawings — 26× larger than the largest prior CAD dataset — with a structure-aware annotation engine and the DPSS baseline.
VecFormer: a line-based representation of vector primitives for panoptic symbol spotting.
4,800 questions, 23 tasks, five cognitive levels — the first hierarchical benchmark of LLM knowledge for architecture, engineering and construction.
MCTS for truss layout under stress, displacement and buckling constraints.
Full list on Google Scholar.