Reproduce “3DCodeEdit: Benchmarking Instruction-Guided 3D Program Editing” 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
Binjie Wang, Xinran Zhao, Shoubin Yu, Xiang Li, Silong Yong, Chenyang Yang, Tianyu Chen, Ruofei Du and 3 more
3DCodeEdit tests models editing existing Blender programs to meet instructions while preserving unchanged parts. It defines all-pass rate as the percentage of tasks whose program runs and renders and passes every checklist item. Across 17 models, the mean is 37.1%, falling to 24.6% for structural edits. On a 312-task subset, GPT-6.1-sol reaches 85.6% with reviewer feedback over eight rounds.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.