Reproduce “Geometry-Aware Test-Time Learning for Quantitative Spatial Reasoning” 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
Gege Zhang, Shuaicheng Niu, Gang Dai, Lei Sun, Shuangping Huang
TTL-SR adapts a vision-language model at test time for estimating 3D distances, directions, and object dimensions from one image and a question. It generates related queries, checks cross-query geometry and semantic agreement, and updates only lora adapters when predictions pass a tolerance filter. On Q-Spatial-ScanNet, Qwen3-VL-4B-Instruct’s five-category mean accuracy rises from 40.00% to 46.47%, using a 25% relative-error threshold.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.