Visual Abstract Thinking Empowers Multimodal Reasoning

Dairu Liu^2*, Ziyue Wang^{1* 🔰}, Minyuan Ruan^1*, Fuwen Luo¹,
Chi Chen¹, Peng Li^{1 ✉️}, Yang Liu^{1 ✉️}

¹Dept. of Computer Science & Technology, Institute for AI, Tsinghua University, Beijing, China
²College of Software, Nankai University, Tianjin, China

^*Equal contribution
^🔰Project lead
^✉️Corresponding author

📢📢📢News

1. [TODO: More interesting cases are on the way]
2. [26-May-2025] Repo and arxiv released

Introduction

Images usually convey richer detail than text, but often include redundant information which potentially downgrades multimodal reasoning performance. When faced with lengthy or complex messages, humans tend to employ abstract thinking to convert them into simple and concise abstracts. Inspired by this cognitive strategy, we introduce Visual Abstract Thinking (VAT), a novel thinking paradigm that prompts Multimodal Large Language Models (MLLMs) with visual abstract instead of explicit verbal thoughts or elaborate guidance, permitting a more concentrated visual reasoning mechanism.

Explicit thinking, such as Chain-of-thought (CoT) or tool-augmented approaches, increases the complexity of reasoning process via inserting verbose intermediate steps, external knowledge or visual information. In contrast, VAT reduces redundant visual information and encourages models to focus their reasoning on more essential visual elements.

Experimental results show that VAT consistently empowers different models, and achieves an average gain of 17% over GPT-4o baseline by employing diverse types of visual abstracts, compared with 13% of CoT, demonstrating that VAT better enhances visual reasoning abilities for MLLMs regarding conceptual, structural and relational reasoning tasks. VAT is also compatible with CoT in knowledge-intensive multimodal reasoning tasks. These findings highlight the effectiveness of visual reasoning via abstract thinking and encourage further exploration of more diverse reasoning paradigms from the perspective of human cognition.

Installation

We recommend using a conda environment:

conda create -n vat python=3.9
conda activate vat

pip install Pillow opencv-python numpy gradio gradio_client networkx scipy datasets

---

Image to Abstract Model Services

The img2abs directory contains two main modules for image abstraction:

1. informative-drawings

• open_server.py: Open-sketch style (port 8086)
• contour_server.py: Contour style (port 8085)
• ani_server.py: Anime style (port 8084)

To start a service (example for open_server.py):

cd img2abs/informative-drawings
python open_server.py

Replace with the corresponding _server.py file for other styles.

2. PhotoSketch

• ps_server.py: Photo-to-sketch style (port 8083)

To start:

cd img2abs/PhotoSketch
python ps_server.py

All services use Gradio and provide a web interface for image upload and abstraction.

---

Data

You can download all task data from Feishu Cloud. After downloading, put all task folders under the ./tasks directory.

The structure should look like this:

./tasks/
├── blink_spatial/
├── blink_counting/
├── blink_viscorr/
├── blink_semcorr/
├── blink_functional_correspondence/
├── mme_commonsense_reasoning/
├── mme_count/
├── mme_existence/
├── mme_position/
├── mme_color/
├── ooo/
└── ...

Running Main Tasks

The run.sh script in the project root can be used to batch run various tasks. It will automatically call scripts in subdirectories such as vat, supporting multiple models and tasks in parallel.

To run:

bash run.sh

The script will launch all configured tasks in the background. Logs and outputs will be saved in the corresponding subdirectories.

Running a Single Task

You can run a single task with the following command:

python run_task.py --task your_task

Replace your_task with the name of the task you want to run.

Environment Variables

• MODEL: Specify the model to use.

  • gpt-4o: gpt-4o-2024-11-20
  • gemini: gemini-2.0-pro-exp-02-05

• ABS_TYPE: Specify the type of visual abstract. Options are:

  • ps: PhotoSketch
  • open: OpenSketch
  • contour: Contour style
  • anime: Anime style

Example:

MODEL=gpt-4o ABS_TYPE=open python run_task.py --task ooo

Supported Tasks and Benchmarks

Our VAT framework supports a variety of tasks for visual perception and multimodal reasoning, covering different benchmarks. Below is a list of all available tasks and their corresponding benchmarks:

Task Name	Benchmark
mme_existence	MME
mme_count	MME
mme_color	MME
mme_position	MME
mme_commonsense_reasoning	MME
blink_counting	BLINK
blink_spatial	BLINK
blink_viscorr	BLINK
blink_semcorr	BLINK
blink_functional_correspondence	BLINK
vd_illusion	HallusionBench
ooo	Odd-One-Out
direction	CoSpace
angle	CoSpace
dif-ang	CoSpace
diff	CoSpace

These tasks cover object-centric perception, object relation reasoning, spatial reasoning, and commonsense reasoning.

Note: For the tasks blink_viscorr, blink_semcorr, and blink_functional_correspondence, you need to set the environment variable ADD_DOT=1 to mark reference points in the visual abstract. For example:

ADD_DOT=1 MODEL=gpt-4o ABS_TYPE=open python run_task.py --task blink_viscorr

Evaluation

You can use calc.py to evaluate the results of your tasks. The basic usage is:

python calc.py <outputs_folder_path> [--accp 0/1] [--abs-type <abstract_type>] [--model <model_name>]

• <outputs_folder_path>: Path to the folder containing the output results to be evaluated.
• --accp 0/1: Whether to calculate the acc+ metric for MME tasks (set to 1 for acc+, 0 for standard accuracy).
• --abs-type <abstract_type>: Specify the type of abstract to evaluate (e.g., ps, open, contour, anime).
• --model <model_name>: Specify the model name to be evaluated (e.g., gpt-4o, gemini).

Example:

python calc.py <outputs_folder_path> --accp 1 --abs-type open --model gpt-4o

This will evaluate all results in the <outputs_folder_path> folder, calculate acc+ for MME tasks, use the open abstract type, and specify the model as gpt-4o.

Note: calc.py can only be used to evaluate tasks other than the CoSpace benchmark.

---

Citation

If you find our work helpful, please cite us:

@misc{liu2025Visual,
  title = {Visual {{Abstract Thinking Empowers Multimodal Reasoning}}},
  author = {Liu, Dairu and Wang, Ziyue and Ruan, Minyuan and Luo, Fuwen and Chen, Chi and Li, Peng and Liu, Yang},
  year = {2025},
  month = may,
  number = {arXiv:2505.20164},
  eprint = {2505.20164},
  primaryclass = {cs},
  publisher = {arXiv},
  doi = {10.48550/arXiv.2505.20164},
  url = {http://arxiv.org/abs/2505.20164},
  archiveprefix = {arXiv},
  langid = {american},
}

---

Acknowledgements

This project integrates and adapts code from the following open-source projects:

• informative-drawings
• PhotoSketch

For further details, please refer to the README files in each submodule.