Reproduce “GaugeVLM: Structuring Spatial Supervision with Measured Geometric Interventions” 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
Hongbo Wang, Zihan Lin, Wenkui Yang, Shiran Ge, Yuang Ai, Jie Cao, Huaibo Huang, Ran He
GaugeVLM trains vision-language models for spatial reasoning using rendered 3D scenes with measured object relations. GaugeDPO sets preference margins from geometric errors, ranks correct relations across camera views, and links answer changes to object interventions. On Qwen2.5-VL-7B, it raises MSMU distance accuracy from 47.5% to 62.5% over GaugeSFT, using a multiplicative error test below 1.25.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.