SafeLens: Prudent and Efficient Video Guardrails via Fast-and-Slow Screening

SafeLens is a video guardrail framework using a fast-and-slow reasoning architecture that allocates different amounts of compute according to input difficulty, achieving efficient and accurate content moderation. Through influence-guided filtering, it retains only 2.4% of the SafeWatch Dataset, and uses structured chain of thought to enable test-time reasoning. It achieves SOTA on real and AI-generated video benchmarks, surpassing open-source guardrails such as SafeWatch-8B and OmniGuard-7B and closed-source models such as GPT-5.4 and Gemini-3.1-pro, while significantly reducing inference cost.