Reproduce “Spatial Latent Reasoning for Embodied Reference Understanding” 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
Ling Li, Jianhui Zhong, Wei Liu, Zheng Jiang aand Yuxuan Liu, Jingyu Li, Zhidong Deng
SLR trains recurrent continuous states to represent fingertip and finger direction, then align four later states with pooled target-region visual features before predicting a box. Finger and region annotations supervise training, not inference. On EgoPoint-Ground, Qwen3.5-4B reaches mIoU 0.778 versus 0.750 for same-backbone SFT, a 2.8-point gain, though gains vary by setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.