README.md

metadata

title: GPD Spatial Reasoning
emoji: 🧭
colorFrom: pink
colorTo: indigo
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Spatial-reasoning VLM distilled with routed 3D privilege
python_version: '3.12'

🧭 GPD · Distilling Routed 3D Privilege for Spatial Reasoning

This Space runs GPD-4B, an RGB-only vision-language model fine-tuned from Qwen3-VL-4B-Instruct with GPD (Geometry-Privileged Distillation): on-policy self-distillation in which a teacher version of the model sees question-routed textual 3D cues (depth / semantic / bird's-eye-view) plus the reference answer, while the deployed student stays RGB-only. The privileged KL on incorrect trajectories augments GRPO.

Upload a single image, ask a 3D spatial-reasoning question — observer-shift spatial imagination, directional relations between objects, depth / distance estimation in meters — and the model thinks step by step, then wraps its final answer in <answer> tags.

Model

• Checkpoint: xinyili0624/GPD-4B (Apache-2.0)
• 2B variant: xinyili0624/GPD-2B
• Paper: Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models
• Code: github.com/ZJU-REAL/GPD
• Training data: xinyili0624/GPD-15k (VSI 10k + SPAR 4k + MindCube 1k)

On Qwen3-VL-4B, GPD reaches 57.1 on VSI-Bench and 37.6 average across MindCube, SPAR-Bench, MMSI-Bench, and ViewSpatial.

Examples

The bundled example images and questions are taken from the authors' GPD-15k dataset (Apache-2.0), specifically its SPAR single-image scenes, which are rendered with bounding-box / point annotations marking the objects the question refers to.

Citation

@article{li2026gpd,
  title={Distilling Routed 3D Privilege for Spatial Reasoning in Vision-Language Models},
  author={Li, Hongxing and Li, Yixin and Li, Dingming and Wang, Zixuan and Yan, Yuchen and Zhang, Wenqi and Lu, Weiming and Shen, Yongliang},
  journal={arXiv preprint arXiv:2610.12355},
  year={2026}
}