File size: 2,892 Bytes
774c8dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5a986e8
2b106b3
 
 
5a986e8
2b106b3
 
 
 
 
 
 
 
 
 
 
 
 
5a986e8
2b106b3
 
 
 
 
 
 
 
5a986e8
774c8dc
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
---
license: cc
task_categories:
- question-answering
language:
- en
tags:
- math
pretty_name: Can Vision-Language Models Solve Visual Math Equations?
size_categories:
- 1K<n<10K
---

# Visual Equation Solving Benchmark

This repository contains the dataset introduced in the paper:  

**Can Vision-Language Models Solve Visual Math Equations?**  which is currently accepted in EMNLP 2025 (Main)

> Despite strong performance in vision and language understanding, Vision-Language Models (VLMs) struggle on tasks requiring integrated perception and symbolic reasoning. This benchmark evaluates VLMs on **visual equation solving**, where systems of linear equations are represented using **object icons** as variables and **icon repetition** as coefficients.  

---

## 📖 Overview

The **Visual Equation Solving Benchmark** tests whether VLMs can:  
1. **Recognize variables** represented as object icons.  
2. **Count coefficients** by inferring from repeated instances of icons.  
3. **Integrate recognition with symbolic reasoning** to solve equations.  

We provide multiple settings:  
- **Symbolic equations** (textual form, rendered as images).  
- **Visual-symbolic equations** (icons for variables, numeric text for coefficients).  
- **Fully visual equations** (both variables and coefficients represented visually).  

---

Example:  
```

🍎🍎🍎 + 🍌🍌 = 10
🍎 + 🍌🍌🍌🍌🍌 = 15

```

---

## 📂 Dataset Structure


There are 2 variants of the dataset based on number of variables used -
2 variables and 3 variables which can be found in the respective zip files.
Once you extract any of them you will see the following tree -
```
├── char_only
│   └── metadata.csv
│   └── *.png
├── counting
│   └── metadata.csv
│   └── *.png
├── icon_only
│   └── metadata.csv
│   └── *.png
├── icon_partial
│   └── metadata.csv
│   └── *.png
└── [two/three]-vars.txt
```
The char_only, icon_only, icon_partial, counting points to the
symbolic, visual, visual-symbolic and counting datasets mentioned in the paper
respectively. Each of them consist of the following metadata -
1. file_path to corresponding image
2. solution to variable
3. mapping to symbolic variable (in case of visual, visual-symbolic, counting dataset)

The base equations which are used to create the same are attached in the respective .txt file in the root level directory.
 
---

## 📜 License

This dataset is released under the **CC BY 4.0 License**.
You are free to share, adapt, and build upon the data with attribution.

---

## 📚 Citation

If you use this dataset, please cite:

```bibtex
@inproceedings{anonymous2025vlm-math,
  title     = {Can Vision-Language Models Solve Visual Math Equations?},
  author    = {Anonymous},
  booktitle = {ACL (under review)},
  year      = {2025}
}
```