REINS steers video AI away from unsafe content without retraining

Open-weight video diffusion models can generate photorealistic unsafe content like violence or misinformation, but existing defenses require expensive fine-tuning that degrades performance or rely on external filters easily bypassed by adversarial prompts. Researchers from UC Riverside introduce REINS (REpresentation-space INference-time Safety steering), a training-free method that realigns video diffusion models at inference time by steering their internal representations toward safe generation. The key insight: safety-relevant structure is linearly encoded in hidden-state activations of video diffusion transformers, and a single direction discovered via Supervised PCA on binary safety labels suffices to separate safe from unsafe generation trajectories. At inference, adding this direction to hidden states at an intermediate transformer layer redirects generation from harmful content to semantically related safe alternatives, with no weight updates, no concept enumeration, and negligible computational overhead.

Mechanistic analysis reveals that while safety information accumulates monotonically with transformer depth, steering effectiveness peaks at intermediate layers (~50% depth), exposing a fundamental tradeoff between information availability and downstream propagation capacity. The researchers evaluated REINS across 9 video diffusion models, multiple parameter scales (1.3B–5B), and both text-to-video and image-to-video generation, calling it the broadest safety evaluation suite in the video generation literature. This approach offers a practical, lightweight solution for deploying safer open-weight video models without sacrificing general capabilities or requiring costly retraining.