GeoReader: Learning Latent Geometry for Spatial Reasoning in Vision-Language Models

Rong Li^1,6* · Shijie Li^2* · Jiangyong Huang^3,6 · Lingdong Kong⁴ · Jiaming Zhou¹
Dongyue Lu⁴ · Zeying Gong¹ · Baoxiong Jia^6†✉ · Junwei Liang^1,5✉ · Siyuan Huang^6✉

¹The Hong Kong University of Science and Technology (Guangzhou)  ·  ²Agency for Science, A*STAR
³Peking University  ·  ⁴National University of Singapore
⁵The Hong Kong University of Science and Technology  ·  ⁶Beijing Institute for General Artificial Intelligence
^*Equal contribution  ·  ^†Project lead  ·  ^✉Corresponding authors

           

 

Geometry foundation models can recover 3D structure from RGB views, but geometric features alone do not ensure that a vision-language model uses them to answer questions. GeoReader learns geometric tokens that the model can read and use for spatial reasoning.

Method

GeoReader first trains Geometric Readers to express predicted geometry through a frozen VLM. It then freezes the Readers and learns question-conditioned geometry injection for spatial QA. Geometry descriptions supervise training; inference uses continuous geometric tokens.

Visualizations

Answer-token attention. Figure 3 compares attention to geometric, RGB, and question tokens.

Spatial reasoning. Figure 4 shows cross-view questions answered with geometry predicted from RGB views; the point clouds come from predicted depth.

More visualizations — supplementary Figures 9–11

Figure 9 · Answer-token attention across question types

Figure 10 · More correct examples

Figure 11 · Failure cases

Training data

Release status

Coming soon.

Citation

@misc{li2026georeader,
  title  = {{GeoReader}: Learning Latent Geometry for Spatial Reasoning in Vision-Language Models},
  author = {Rong Li and Shijie Li and Jiangyong Huang and Lingdong Kong and Jiaming Zhou and Dongyue Lu and Zeying Gong and Baoxiong Jia and Junwei Liang and Siyuan Huang},
  year   = {2026},
  eprint = {2610.18694},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url = {https://arxiv.org/abs/2610.18694}
}