HarmVideoBench: Benchmarking Harmful Video Understanding in Large Multimodal Models
Jiajun Wu, Haoyu Kang, Yining Sun, Jiacheng Hou, Heng Zhang, Danyang Zhang, Zhenjun Zhao, Haochi Zhang, Leixin Sun, Eric Hanchen Jiang, Yushan Li, Ruiyu Li, Mengkai Huang, Yan Gao, Xu Zhang, Guancheng Wan
Read on arXiv →
Key claim
BCR improves harmful video understanding accuracy to 84.4%.
In plain English
Imagine you're trying to build a system that can automatically identify harmful videos online. The challenge is that harmful content isn't always obvious; it can be subtle and layered, making it hard for models to detect. Currently, many systems just classify videos as harmful or not, which misses the deeper context and nuances of what makes a video harmful. This is what's called binary classification failure — it oversimplifies a complex problem and can lead to misjudgments about content that might actually be harmful in a more nuanced way.
To tackle this, the authors created HarmVideoBench, a new benchmark that goes beyond simple yes-or-no answers. Instead of just asking if a video is harmful, it evaluates models on three different levels: what can be seen in the video, the meaning behind the clips, and the reasoning that goes beyond just the clips themselves. This multi-layered approach helps capture the complexities of harmful content better than previous methods.
They also introduced a method called BCR, which helps models understand the context of harmful videos more effectively. By dynamically retrieving relevant information only when needed, BCR significantly boosts the performance of models on this new benchmark, raising their accuracy from 61.7% to an impressive 84.4%. For anyone building systems for content moderation, this means you can now evaluate and improve your models in a way that truly reflects their understanding of harmful content, rather than relying on surface-level assessments.
The introduction of a multi-layered benchmark for harmful video evaluation significantly extends existing methodologies.
The evaluation of 19 leading models on a well-defined benchmark provides solid empirical support for the claims made.
Deep reliability assessment
The methodology supports HarmVideoBench as a diagnostic MCQ benchmark for layered harmful-video understanding, especially distinguishing surface evidence, clip-internal meaning, and beyond-clip reasoning. The stronger claim that BCR meaningfully advances real-world moderation is overclaimed, because its retrieval bank is benchmark-aligned and the benchmark mostly lacks benign false-positive cases.
Reproducibility
Partial. The paper describes a 1,379-video / 4,137-question benchmark and says the benchmark is released as a research resource, but the code repository is listed only as 'Coming soon' and no concrete repository URL is provided.
Key figure
Figure 1 contrasts conventional binary harmful-video evaluation with HarmVideoBench's three-level evaluation framework, where models answer multidimensional questions about observable evidence, internal clip meaning, and beyond-clip reasoning.