Research map / Video / Video-language understanding
SafeWatch: An Efficient Safety-Policy Following Video Guardrail Model with Transparent Explanations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SafeWatch, an InternVL2-8B-based video guardrail that encodes safety policies in parallel with equivalent RoPE positions and prunes video tokens per policy, plus a 2M-video benchmark (SafeWatch-Bench) annotated by a multi-agent propose-discuss pipeline.
Problem
Video guardrail models are either simple classifiers with fixed categories and no explanations, or MLLM prompting approaches that are slow, suffer policy position bias, give vague explanations, and adapt poorly to new policies; existing video guardrail benchmarks are small and binary-labeled.
Approach
Encode each safety policy chunk in parallel with masked attention and equivalent RoPE position embeddings to remove positional bias, and prune video tokens per policy using policy-video cross-attention relevance scores, trained in three stages (multi-task guardrail SFT, adaptive-pruning SFT, preference post-tuning).
Key results
- On SafeWatch-Bench-Real, SafeWatch-8B reaches 72.6% average accuracy and 86.7 F1 versus 43.4% accuracy for GPT-4o, and cuts inference time by 0.4 s versus InternVL2-8B.
- Pruning up to 90% of video tokens keeps performance within 1% drop, per the paper's pruning-ratio analysis.
- On three unseen policies (children, firearms, accidents), SafeWatch-8B averages 81.8/87.8/78.5 accuracy, above GPT-4o on children and accidents.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add SafeWatch training and modules to InternVL2-8B | InternVL2-8B · SafeWatch-Bench-Real | accuracy: 29.1 -> 72.6 | Helped |
| add PEPE (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 65.2 -> 72.6 | Helped |
| add PAP (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 69.9 -> 72.6 | Helped |
| add PAP (ablation) | SafeWatch-8B · SafeWatch-Bench | throughput time (s): 4.2 -> 3.9 | Helped |
| add DPO preference post-tuning (ablation) | SafeWatch-8B · SafeWatch-Bench | explanation rating: 6.12 -> 7.17 | Helped |
| set pruning ratio to 20% instead of the full method setting (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 72.6 -> 71.6 | Hurt |
| set pruning ratio to 40% instead of the full method setting (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 72.6 -> 72.4 | Hurt |
| set pruning ratio to 95% instead of the full method setting (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 72.6 -> 65.3 | Hurt |
| set pruning ratio to 99% instead of the full method setting (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 72.6 -> 55.9 | Hurt |
| include explicit policy definitions during SFT (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 52.3 -> 62.0 | Helped |
| advance from Stage-1 to Stage-2 adaptive-pruning training (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 63.9 -> 67.3 | Helped |
| advance from Stage-2 to Stage-3 preference post-tuning (ablation) | SafeWatch-8B · SafeWatch-Bench | accuracy: 67.3 -> 72.6 | Helped |
Surprising findings
- Pruning up to 90% of video tokens keeps accuracy within 1% of the unpruned model.
- GPT-4o rates SafeWatch's explanations higher than its own by a 10% margin, despite the common self-preference bias.
Limitations
- Generative subset performance is about 18% lower than real-world, attributed to generator limits in abuse-related categories.
- Inference speed comparisons with non-explanatory models are not fully fair due to differing parameter scales and response lengths.
- Relies on GPT-4o as judge and as the consensus judge in annotation pipeline.
Experimental setup
Datasets: SafeWatch-Bench LSPD XD-Violence UCF-Crime FakeSV FVC MoB
Benchmarks: SafeWatch-Bench-Real SafeWatch-Bench-GenAI LSPD XD-Violence UCF-Crime FakeSV FVC MoB children safety
Metrics: accuracy F1 AUPRC explanation rating (0-10) inference time
Baselines: GPT-4o Gemini-1.5-pro InternVL2-8B InternVL2-26B LlavaGuard-34B Holmes-VAD LlamaGuard3V-11B Azure Mod API
Builds on: InternVL2 TransnetV2 AutoShot DPO RoPE
Closest papers by meaning
- Towards Policy-Adaptive Image Guardrail: Benchmark and Method (CVPR 2026 · similarity 0.93)
- PolyGuard: Massive Multi-Domain Safety Policy-Grounded Guardrail Dataset (NeurIPS 2025 · similarity 0.88)
- Controllable Multi-label Video Safety Detection via Adaptive Tversky Policy Optimization (NeurIPS 2026 · similarity 0.87)
- Video-SafetyBench: A Benchmark for Safety Evaluation of Video LVLMs (NeurIPS 2025 · similarity 0.87)
- SafeVid: Toward Safety Aligned Video Large Multimodal Models (NeurIPS 2025 · similarity 0.87)
- GuardTrace-VL: Detecting Unsafe Multimodel Reasoning via Iterative Safety Supervision (CVPR 2026 · 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.