TL; DR: We propose TreeBench, the first benchmark specially designed for evaluating "thinking with images" capabilities with traceable visual evidence, and TreeVGR, the current state-of-the-art open-source visual grounded reasoning models.
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 visual grounded reasoning.
- [2025/07/12] 🔥🔥🔥 TreeBench and TreeVGR have been supported by VLMEvalKit! 🔥🔥🔥
- [2025/07/11] 🔥 TreeBench and TreeVGR have been released. Check out the paper for details.
pip3 install -r requirements.txt
pip3 install flash-attn --no-build-isolation -vThis 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 TreeVGRand then, simply run inference_treebench.py
python3 inference_treebench.pyThis 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 utlized VLMEvalKit for a more comprehensive evaluation.
Benchmark
Checkpoints
- TreeVGR-7B: https://huggingface.co/HaochenWang/TreeVGR-7B
- TreeVGR-7B-CI: https://huggingface.co/HaochenWang/TreeVGR-7B-CI
Training Datasets
- TreeVGR-RL-37K: https://huggingface.co/datasets/HaochenWang/TreeVGR-RL-37K
- TreeVGR-SFT-35K: https://huggingface.co/datasets/HaochenWang/TreeVGR-SFT-35K
If you find Ross 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}
}We would like to express our sincere appreciation to the following projects:
- Qwen2.5-VL: The base model we utilzed.
- 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.