Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -31,3 +31,55 @@ configs:
|
|
31 |
- split: train
|
32 |
path: data/train-*
|
33 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
- split: train
|
32 |
path: data/train-*
|
33 |
---
|
34 |
+
|
35 |
+
# VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge
|
36 |
+
|
37 |
+
[๐ Homepage](https://neulab.github.io/VisualPuzzles/) | [๐ VisualPuzzles](https://huggingface.co/datasets/neulab/VisualPuzzles) | [๐ป Github](https://github.com/neulab/VisualPuzzles) | [๐ Arxiv](https://arxiv.org/abs/2504.10342) | [๐ PDF](https://arxiv.org/pdf/2504.10342) | [๐ฅ๏ธ Zeno Model Output](https://hub.zenoml.com/project/2e727b03-a677-451a-b714-f2c07ad2b49f/VisualPuzzles)
|
38 |
+
|
39 |
+
## Overview
|
40 |
+
**VisualPuzzles** is a multimodal benchmark specifically designed to evaluate **reasoning abilities** in large models while deliberately minimizing reliance on domain-specific knowledge.
|
41 |
+
|
42 |
+
Key features:
|
43 |
+
- 1168 diverse puzzles
|
44 |
+
- 5 reasoning categories: Algorithmic, Analogical, Deductive, Inductive, Spatial
|
45 |
+
- Difficulty labels: Easy, Medium, Hard
|
46 |
+
- Less knowledge-intensive than existing benchmarks (e.g., MMMU)
|
47 |
+
- More reasoning-complex than existing benchmarks (e.g., MMMU)
|
48 |
+
|
49 |
+
## Key Findings
|
50 |
+
- All models perform worse than humans; most can't surpass even 5th-percentile human performance.
|
51 |
+
- Strong performance on knowledge-heavy benchmarks does not transfer well.
|
52 |
+
- Larger models and structured "thinking modes" don't guarantee better results.
|
53 |
+
- Scaling model size does not ensure stronger reasoning
|
54 |
+
|
55 |
+
## Usage
|
56 |
+
|
57 |
+
To load this dataset via Hugging Faceโs `datasets` library:
|
58 |
+
|
59 |
+
```python
|
60 |
+
from datasets import load_dataset
|
61 |
+
|
62 |
+
dataset = load_dataset("neulab/VisualPuzzles")
|
63 |
+
data = dataset["train"]
|
64 |
+
|
65 |
+
sample = data[0]
|
66 |
+
print("ID:", sample["id"])
|
67 |
+
print("Category:", sample["category"])
|
68 |
+
print("Question:", sample["question"])
|
69 |
+
print("Options:", sample["options"])
|
70 |
+
print("Answer:", sample["answer"])
|
71 |
+
```
|
72 |
+
|
73 |
+
## Citation
|
74 |
+
|
75 |
+
If you use or reference this dataset in your work, please cite:
|
76 |
+
|
77 |
+
```bibtex
|
78 |
+
@article{song2025visualpuzzles,
|
79 |
+
title = {VisualPuzzles: Decoupling Multimodal Reasoning Evaluation from Domain Knowledge},
|
80 |
+
author = {Song, Yueqi and Ou, Tianyue and Kong, Yibo and Li, Zecheng and Neubig, Graham and Yue, Xiang},
|
81 |
+
year = {2025},
|
82 |
+
journal = {arXiv preprint arXiv:2504.10342},
|
83 |
+
url = {https://arxiv.org/abs/2504.10342}
|
84 |
+
}
|
85 |
+
```
|