README.md

---
dataset_info:
  features:
  - name: group_id
    dtype: string
  - name: mvh_type
    dtype: string
  - name: category
    dtype: string
  - name: view_1
    dtype: image
  - name: view_2
    dtype: image
  - name: mc_questions
    list: string
  - name: mc_options
    list: string
  - name: mc_answers
    list: string
  - name: binary_questions
    list: string
  - name: binary_answers
    list: string
  splits:
  - name: validation
    num_bytes: 23895547
    num_examples: 80
  - name: test
    num_bytes: 196420891
    num_examples: 720
  download_size: 219687633
  dataset_size: 220316438
configs:
- config_name: default
  data_files:
  - split: validation
    path: data/validation-*
  - split: test
    path: data/test-*
task_categories:
- visual-question-answering
language:
- en
tags:
- multimodal,
- multi-view,
- hallucination,
- multi-image,
- egocentric,
- exocentric,
- visual-question-answering
pretty_name: MVH-Bench
---

# MVH-Bench

**MVH-Bench** is the benchmark introduced in:

**Revealing Multi-View Hallucination in Large Vision-Language Models**  
Wooje Park<sup>\*</sup>, Insu Lee<sup>\*</sup>, Soohyun Kim, Jaeyun Jang, Minyoung Noh, Kyuhong Shim, and Byonghyo Shim (* Equal contribution)  
**EMNLP 2026 Main**

[**Paper**](https://arxiv.org/abs/2603.23934)

MVH-Bench is a multi-view visual question answering benchmark designed to evaluate **Multi-View Hallucination (MVH)** in Large Vision-Language Models (LVLMs).

The dataset contains **4,800 question-answer pairs** organized into **800 image-pair groups**, with each group containing two multiple-choice questions and four binary questions.

| Split | Image-Pair Groups | MC QA | Binary QA | Total QA |
|---|---:|---:|---:|---:|
| Validation | 80 | 160 | 320 | 480 |
| Test | 720 | 1,440 | 2,880 | 4,320 |
| **Total** | **800** | **1,600** | **3,200** | **4,800** |

## Categories

MVH-Bench contains two types of multi-view hallucination:

- `cross_view`: hallucination caused by visual evidence from an incorrect viewpoint
- `cross_instance`: hallucination caused by visual evidence from an incorrect instance

The benchmark contains four evaluation categories:

- Action
- Object
- Numerical
- Spatial

Each image-pair group contains two question formats:

- Multiple Choice (MC)
- Binary (Yes/No)

## Dataset Structure

Each example corresponds to one image-pair group and contains the following fields:

- `group_id`: unique group identifier
- `mvh_type`: MVH type (`cross_view` or `cross_instance`)
- `category`: question category
- `view_1`: first image in the multi-view pair
- `view_2`: second image in the multi-view pair
- `mc_questions`: two multiple-choice questions
- `mc_options`: three answer choices shared by the two MC questions
- `mc_answers`: ground-truth answers for the two MC questions
- `binary_questions`: four binary questions
- `binary_answers`: ground-truth answers for the four binary questions


## Usage

```python
from datasets import load_dataset

dataset = load_dataset(
    "SNU-ISLAB/MVH-Bench",
    split="test",
)

sample = dataset[0]

print(sample["mvh_type"])
print(sample["category"])

print(sample["mc_questions"])
print(sample["mc_options"])
print(sample["mc_answers"])

print(sample["binary_questions"])
print(sample["binary_answers"])

view_1 = sample["view_1"]
view_2 = sample["view_2"]
```

The validation split can be loaded with:

```python
validation = load_dataset(
    "SNU-ISLAB/MVH-Bench",
    split="validation",
)
```

## Evaluation


For multiple-choice questions, we report:

- **Accuracy (Acc)**: accuracy over individual MC questions
- **Pair Accuracy (p-Acc)**: the proportion of image-pair groups in which both MC questions are answered correctly
- **Adversarial Error Rate (AER)**: among incorrect MC predictions, the proportion that select the adversarial alternative rather than the `Neither` option

For binary questions, we report:

- **Accuracy (Acc)**: accuracy over individual Yes/No questions
- **Pair Accuracy (p-Acc)**: the proportion of positive-negative question pairs in which both questions are answered correctly
- **Quad Accuracy (q-Acc)**: the proportion of image-pair groups in which all four binary questions are answered correctly
- **Yes Error Rate (YER)**: the proportion of erroneous binary predictions biased toward `Yes`

The overall **MVH-Score** jointly measures performance on the multiple-choice and binary evaluations:  
**MVH-Score = Binary Acc + Binary p-Acc + Binary q-Acc + MC Acc + MC p-Acc**

AER and YER are reported as diagnostic error metrics and are not included in the MVH-Score.


Standardized evaluation for **MVH-Bench** is supported in [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval).

### Example

```bash
python -m lmms_eval \
  --model qwen2_5_vl \
  --model_args pretrained=Qwen/Qwen2.5-VL-7B-Instruct \
  --tasks mvh_bench \
  --batch_size 1
```

### Available Tasks

- `mvh_bench`: Full benchmark on the test split, including both cross-instance and cross-view evaluations
- `mvh_bench_val`: Validation split
- `mvh_bench_cross_instance`: Cross-instance subset of the test split
- `mvh_bench_cross_view`: Cross-view subset of the test split


## Citation

If you use MVH-Bench in your research, please cite:

```bibtex
@article{park2026revealing,
  title={Revealing Multi-View Hallucination in Large Vision-Language Models},
  author={Park, Wooje and Lee, Insu and Kim, Soohyun and Jang, Jaeyun and Noh, Minyoung and Shim, Kyuhong and Shim, Byonghyo},
  journal={arXiv preprint arXiv:2603.23934},
  year={2026}
}
```