3DThinker-Mindcube

This repository contains the stage 1 model checkpoint for 3DThinker, as presented in the paper Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views.

3DThinker is a framework that enables Vision-Language Models (VLMs) to exploit geometric information within images for 3D spatial reasoning, simulating human-like spatial imagination without requiring explicit 3D prior inputs or labeled 3D training data.

• Paper: Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views
• Code: GitHub - zhangquanchen/3DThinker

Introduction

• The model was trained on Mindcube_Train and tested on MindCube-Tiny.
• This model corresponds to stage 1 training (supervised alignment of 3D latents) of Qwen2.5-3B-VL.
• Note that Tab. 2 in the paper is trained on a different training data configuration.

Bibtex

If you find 3DThinker helpful for your work, please cite:

@article{chen2025think,
  title={Think with 3D: Geometric Imagination Grounded Spatial Reasoning from Limited Views},
  author={Chen, Zhangquan and Zhang, Manyuan and Yu, Xinlei and Luo, Xufang and Sun, Mingze and Pan, Zihao and Feng, Yan and Pei, Peng and Cai, Xunliang and Huang, Ruqi},
  journal={arXiv preprint arXiv:2510.18632},
  year={2025}
}