Reproduce “SpatialMind: Adaptive Multi-Level Geometric Fusion for Spatial Reasoning in Multimodal Large Language Models” 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
Shibin Zhao, Hao Tan, Fan Yang, Haixin Wang, Feilong Chen, Yidong Ma, Chaoyang Zhao, Ming Tang and 1 more
SpatialMind adaptively mixes shallow, intermediate, and deep geometry features for each text-and-scene context. It targets differences between local perception and broader spatial reasoning. Controlled removals show shallow geometry matters more for low-level tasks, while deep geometry matters more for high-level reasoning. SpatialMind-5B scored 71.5 on VSI-Bench, versus 69.6 for training-matched SpatialStack-5B, within the 3–5B group.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.