Research map / Video / Video-language understanding
From Evaluation to Defense: Advancing Safety in Video Large Language Models
Yiwei Sun, Peiqi Jiang, Chuanbin Liu, Luohao Lin, Zhiying Lu, Hongtao Xie
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 |
|---|---|---|---|
| apply VideoSafety-R1 post-training | VideoLLaMA3-2B · VSE-HH | DSR: 18.4% -> 89.5% | Improved |
| apply VideoSafety-R1 post-training | Qwen2.5 VL-7B · VSE-HH | DSR: 57.3% -> 95.4% | Improved |
| apply VideoSafety-R1 post-training | VideoLLaMA3-2B · VSE-SafeQ | FRR: 0.7% -> 13.4% | Hurt |
| apply VideoSafety-R1 post-training | VideoLLaMA3-2B · MVBench | accuracy: 65.3% -> 64.7% | Hurt |
| add visual-attack videos embedding harmful queries | VideoSafety-R1 · VSE-HH-VA | DSR: 89.5% -> 95.7% | No clear effect |
| add alarm token to fine-tuning (ablation) | VideoLLaMA3-2B · VSE-HH | DSR: 74.0% -> 78.1% | Improved |
| add alarm-token classification task (ablation) | VideoLLaMA3-2B · VSE-HH | DSR: 78.1% -> 83.1% | Improved |
| add safety-guided GRPO on top of AT-SFT (ablation) | VideoLLaMA3-2B · MMBench | DSR: 40.3% -> 85.8% | Improved |
| add safety-guided GRPO on top of AT-SFT (ablation) | VideoLLaMA3-2B · VSE-HH | DSR: 83.1% -> 81.5% | Hurt |
| add Dynamic Reward Adaptation (ablation) | VideoSafety-R1 · VSE-HH | DSR: 81.5% -> 89.5% | Improved |
| add video modality to Video LLMs | 21 Video LLMs · VSE-HH | DSR: average 34.2% drop | Hurt |
| increase FPS (ablation) | Qwen2.5 VL · VSE-HH | DSR | Hurt |
Closest papers by meaning
- Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs (NeurIPS 2025 · similarity 0.94)
- Rethinking Bottlenecks in Safety Fine-Tuning of Vision Language Models (ICLR 2026 · similarity 0.92)
- SafeVid: Toward Safety Aligned Video Large Multimodal Models (NeurIPS 2025 · similarity 0.92)
- Risk Awareness Injection: Calibrating Vision-Language Models for Safety without Compromising Utility (ICML 2026 · similarity 0.91)
- Evolving Contextual Safety in Multi-Modal Large Language Models via Inference-Time Self-Reflective Memory (CVPR 2026 · similarity 0.90)
- VLSU: Mapping the Limits of Joint Multimodal Understanding for AI Safety (ICLR 2026 · similarity 0.89)
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.
