Reproduce “PhysMoE: Fast Spatial Grounding with Single-Token Coordinate Representation in Modality-Decoupled VLMs” 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
Ruimeng Liu, Denan Liang, Ruofei Bai, Fuming Zhang, Kuan Xu, Shenghai Yuan, Lihua Xie
PhysMoE treats coordinate pairs as single spatial tokens and routes them to spatial experts while sharing attention with language tokens. It trains pointing with distance fields that allow valid regions rather than one center. Across nine grounding benchmarks, PhysMoE-9B scores 74.2 versus 64.1 for its initialization baseline; coordinate throughput is 23.34 samples per second, not a measure of robot-control speed.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.