TreeVGR-7B-CI

TreeVGR-7B: Traceable Evidence Enhanced Visual Grounded Reasoning Model

This repository contains the TreeVGR-7B model, a state-of-the-art open-source visual grounded reasoning model, as presented in the paper Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodology.

Project page: https://haochenwang409.github.io/TreeVGR/

<p align="center"> <a href="https://huggingface.co/papers/2507.07999"> <img src="https://img.shields.io/badge/Paper-HuggingFace-red"></a> <a href="https://huggingface.co/datasets/HaochenWang/TreeBench"> <img src="https://img.shields.io/badge/TreeBench-HuggingFace-orange"></a> <a href="https://huggingface.co/HaochenWang/TreeVGR-7B"> <img src="https://img.shields.io/badge/TreeVGR-HuggingFace-yellow"></a> <a href="https://github.com/Haochen-Wang409/TreeVGR"> <img src="https://img.shields.io/badge/Code-GitHub-blue"></a> </p>

Abstract

Models like OpenAI-o3 pioneer visual grounded reasoning by dynamically referencing visual regions, just like human "thinking with images". However, no benchmark exists to evaluate these capabilities holistically. To bridge this gap, we propose TreeBench (Traceable Evidence Evaluation Benchmark), a diagnostic benchmark built on three principles: (1) focused visual perception of subtle targets in complex scenes, (2) traceable evidence via bounding box evaluation, and (3) second-order reasoning to test object interactions and spatial hierarchies beyond simple object localization. Prioritizing images with dense objects, we initially sample 1K high-quality images from SA-1B, and incorporate eight LMM experts to manually annotate questions, candidate options, and answers for each image. After three stages of quality control, TreeBench consists of 405 challenging visual question-answering pairs, even the most advanced models struggle with this benchmark, where none of them reach 60% accuracy, e.g., OpenAI-o3 scores only 54.87. Furthermore, we introduce TreeVGR (Traceable Evidence Enhanced Visual Grounded Reasoning), a training paradigm to supervise localization and reasoning jointly with reinforcement learning, enabling accurate localizations and explainable reasoning pathways. Initialized from Qwen2.5-VL-7B, it improves V* Bench (+16.8), MME-RealWorld (+12.6), and TreeBench (+13.4), proving traceability is key to advancing vision-grounded reasoning.

News

• [2025/07/11] 🔥🔥🔥 TreeBench and TreeVGR have been supported by VLMEvalKit! 🔥🔥🔥
• [2025/07/11] 🔥 TreeBench and TreeVGR have been released.

Installation

pip3 install -r requirements.txt
pip3 install flash-attn --no-build-isolation -v

Usage

This repo provides a simple local inference demo of our TreeVGR on TreeBench. First, clone this repo,

git clone https://github.com/Haochen-Wang409/TreeVGR
cd TreeVGR

and then, simply run inference_treebench.py

python3 inference_treebench.py

This should give:

Perception/Attributes 18/29=62.07
Perception/Material 7/13=53.85
Perception/Physical State 19/23=82.61
Perception/Object Retrieval 10/16=62.5
Perception/OCR 42/68=61.76
Reasoning/Perspective Transform 19/85=22.35
Reasoning/Ordering 20/57=35.09
Reasoning/Contact and Occlusion 25/41=60.98
Reasoning/Spatial Containment 20/29=68.97
Reasoning/Comparison 20/44=45.45
==> Overall 200/405=49.38
==> Mean IoU: 43.3

This result is slightly different from the paper, as we mainly utilized VLMEvalKit for a more comprehensive evaluation.

Hugging Face Resources

Benchmark

• TreeBench

Checkpoints

• TreeVGR-7B
• TreeVGR-7B-CI

Training Datasets

• TreeVGR-RL-37K
• TreeVGR-SFT-35K

Citation

If you find this work useful for your research and applications, please cite using this BibTeX:

@article{wang2025traceable,
  title={Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodology},
  author={Haochen Wang and Xiangtai Li and Zilong Huang and Anran Wang and Jiacong Wang and Tao Zhang and Jiani Zheng and Sule Bai and Zijian Kang and Jiashi Feng and Zhuochen Wang and Zhaoxiang Zhang},
  journal={arXiv preprint arXiv:2507.07999},
  year={2025}
}

Acknowledgement

We would like to express our sincere appreciation to the following projects:

• Qwen2.5-VL: The base model we utilized.
• VGR: The source of our SFT dataset.
• V* and VisDrone: The image source of our RL dataset.
• SA-1B: The image source of our TreeBench.
• LLaMA-Factory: The SFT codebase we utilized.
• EasyR1: The RL codebase we utilized.