---
license: apache-2.0
task_categories:
- visual-question-answering
language:
- en
---
# VideoHallu: Evaluating and Mitigating Multi-modal Hallucinations for Synthetic Videos
[Zongxia Li*](https://zli12321.github.io/), [Xiyang Wu*](https://wuxiyang1996.github.io/), [Yubin Qin](https://www.linkedin.com/in/yubin-qin/), [Hongyang Du](https://smashedpython.github.io/HongyangDu.github.io/), [Guangyao Shi](https://guangyaoshi.github.io/), [Dinesh Manocha](https://www.cs.umd.edu/people/dmanocha), [Tianyi Zhou](https://tianyizhou.github.io/), [Jordan Lee Boyd-Graber](https://users.umiacs.umd.edu/~ying/)
[[π Paper](https://arxiv.org/abs/2505.01481)] [[π€ Dataset](https://huggingface.co/datasets/IntelligenceLab/VideoHallu)] [[πWebsite](https://wuxiyang1996.github.io/videohallu_page/)]
## π About VideoHallu
Synthetic video generation using foundation models has gained significant attention due to its realism and broad applications. However, while these models excel at generating visually coherent and high-quality video frames, they often overlook commonsense reasoning and physical law violations, leading to abnormal content. Existing score-based evaluations like [VideoScore](https://arxiv.org/abs/2406.15252) mainly focus on general video quality and do not take these abnormalities into account, and offer no explanations of the evaluation results. A more promising evaluation approach is to leverage multi-modal large language models (MLLMs) as interpretable video evaluators, following the approach of [FActScore](https://arxiv.org/abs/2305.14251). However, how well MLLMs can detect these abnormalities in synthetic videos is underexplored.
Motivated by a more interpretable video generation evaluation, we introduce VideoHallu, a benchmark built from synthetic videos produced by popular models like [Sora](https://openai.com/sora/), [Veo2](https://veo2.ai), [Kling](https://www.klingai.com/global/), paired with expert-crafted question-answering pair examples easily solvable with human-level perception and reasoning across multiple categories. We evaluate several State-of-the-Art (SoTA) MLLMs with our benchmark, including [GPT-4o](https://openai.com/index/hello-gpt-4o/), [Gemini-2.5-Pro](https://deepmind.google/technologies/gemini/pro/), [Qwen-2.5-VL](https://github.com/QwenLM/Qwen2.5-VL), and forefront models like [Video-R1](https://github.com/tulerfeng/Video-R1) and [VideoChat-R1](https://github.com/OpenGVLab/VideoChat-R1). Despite the strong performance of R1 MLLMs on real-world video benchmarks like [MVBench](https://huggingface.co/datasets/OpenGVLab/MVBench) and [MovieChat](https://github.com/rese1f/MovieChat), these models still struggle and hallucinate on basic commonsense and physics reasoning tasks in synthetic videos, highlighting synthetic video hallucination as an underexplored challenge.
Moreover, we post-train current SoTA MLLMs, [Qwen-2.5-VL-7B](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct), with [Group Relative Policy Optimization (GRPO)](https://arxiv.org/abs/2501.12948) using both real-world and synthetic commonsense/physics datasets. Our results show improved overall accuracy compared to the base model, achieving the highest performance among all models, highlighting the importance of integrating high-quality counterexamples to enhance commonsense and physics reasoning in MLLMs' language priors.
## π₯ News
- [2025/05/02] We expand our dataset with more QA pairsπ€.
- [2025/05/02] We release our [datasets](https://huggingface.co/datasets/IntelligenceLab/VideoHallu)π€.
- [2025/05/02] We release our GRPO free-form [RewardModel](https://huggingface.co/IntelligenceLab/RewardPreferenceBert/settings)π€.
## Table of Contents
* [Benchmark](#benchmark)
* [Getting Started](#setup)
* [The Dawn of MLLMs in Synthetic Videos](#showcase)
* [Evaluation over SoTA MLLMs](#evaluation)
* [Reward Model](#rb)
* [Training](#training)
* [Fine-tuning Results](#evaluation_ft)
* [Acknowledgements](#ak)
* [Citations](#citations)
## π Benchmark
We design our benchmark, VideoHallu, around four question categories aimed at probing hallucinations in synthetic video understanding, organized by the level of reasoning required from MLLMs to perform video-question answering in practice. The benchmark spans from perceptual understanding to high-level abstract reasoning.
* **Alignment** checks if the model correctly identifies and understands entities using visual and textual cues.
* **Spatial-temporal Consistency** examines whether the model can track entity motion across frames.
* **Common Sense Reasoning** tests if the model can reason based on its knowledge.
* **Physics** assesses if the model applies physical laws to entity motions and procedural understanding.
Each question in a category may also be assigned to multiple sub-categories, depending on the specific aspects it targets. Detailed annotations and sub-category breakdowns are available [here](https://huggingface.co/datasets/zli12321/VideoHalluB):
| Updated on | HuggingFace | Dataset Size |
|-------------|:------------------------------------------------:|:------------:|
| May, 2, 2025 | [HuggingFace](https://huggingface.co/datasets/zli12321/VideoHalluB) | 3233 |
Below is an overview of our benchmarkβs organization, including the high-level question categories, ranked by the level of reasoning required by MLLMs, and their corresponding sub-category breakdowns.
## π Getting Started
To set up our benchmark, please follow the steps provided below:
```
# Download the synthetic dataset
pip install huggingface_hub
# Download data to your local dir
huggingface-cli download IntelligenceLab/VideoHallu --repo-type dataset --local-dir ./new_video_folders --local-dir-use-symlinks False
# Download and unzip the physben training data videos
curl -L -o video.part1.rar https://huggingface.co/datasets/WeiChow/PhysBench-train/resolve/main/video.part1.rar
# Unzip data (linux system)
unrar x video.part1.rar
```
## π§ The Dawn of MLLMs in Synthetic Videos
We present selected cases from SoTA MLLM evaluations across each category. Hallucinations in model answers, common sense or physics violations in videos, and other notable cues in the video, questions, or ground truth are highlighted to assist the reader's understanding. More examples can be found in the Appendix of [our paper](https://arxiv.org/abs/2505.01481).
**Note:** The legend below explains all the symbols used to represent the State-of-the-Art (SoTA) MLLMs featured in our showcases for synthetic video generation and video question-answering.