Research map / Video / Video-language understanding
Failures to Surface Harmful Contents in Video Large Language Models
Yuxin Cao, Wei Song, Derui Wang, Jingling Xue, Jin Song Dong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| raise PiP scaling ratio eta from 0.2 to 0.5 (ablation) | PPA on LLaVA-Video-7B-Qwen2 · violence, crime, pornography (50 videos) | HOR: HOR below 20% | Improved |
| increase PiP scaling ratio eta up to 0.30 (ablation) | PPA on VideoLLaMA2 and ShareGPT4Video · harmful corner patch | HOR: unresponsive even at eta = 0.3 | No clear effect |
| increase PiP scaling ratio eta (ablation) | PPA on LLaVA-based models · harmful corner patch | HOR: HOR gradually decreases | Improved |
| raise overlay opacity alpha from 0.3 to 0.7 (ablation) | TOA on five VideoLLMs · transparent overlay | HOR: worst-case HOR remains as high as 42 | No clear effect |
| raise overlay opacity alpha (ablation) | TOA on L-7B and LN-32B · transparent overlay | HOR: slight downward trend | Improved |
| use key frame selection instead of uniform sampling (ablation) | FRA on ShareGPT4Video · frame replacement | HOR: 98 average HOR | No clear effect |
| add denser sampling, relevance-based sampling or VLM-assisted detection as mitigation (ablation) | VideoLLMs under attack · omission attacks | HOR: HOR remaining as high as 71% - 95% | Mixed |
| use a more informative prompt such as describing violent scenes (ablation) | VideoLLMs under attack · omission attacks | HOR | No clear effect |
| increase number of uniformly sampled frames to 16 (ablation) | FRA clip duration · 2000-frame simulation | capture by sampled frames: segment shorter than 6% captured by at most one frame | Hurt |
| make harmful content visible via the full VideoLLM instead of the SigLIP encoder alone (ablation) | LLaVA-Video-7B-Qwen2 · harmful video detection (100 per category) | detection accuracy | Hurt |
Closest papers by meaning
- Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs (NeurIPS 2025 · similarity 0.89)
- Protecting Your Video Content: Disrupting Automated Video-based LLM Annotations (CVPR 2025 · similarity 0.87)
- Vid-SME: Membership Inference Attacks against Large Video Understanding Models (NeurIPS 2025 · similarity 0.87)
- VISTA-LLAMA: Reducing Hallucination in Video Language Models via Equal Distance to Visual Tokens (CVPR 2024 · similarity 0.86)
- Distilling Vision-Language Models on Millions of Videos (CVPR 2024 · similarity 0.86)
- VMDT: Decoding the Trustworthiness of Video Foundation Models (NeurIPS 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
