MATH · IN · MODELS

An SVD contrastive subspace enables closed-loop control of video diffusion

measured in 1 paper

Hong et al. propose LA-LQR for text-to-video diffusion steering: per (layer, timestep) they extract an orthonormal basis of a contrastive prompt-pair difference matrix via streaming randomized SVD, finding a 64-dimensional latent captures most contrastive-difference energy [hong-etal-2026-la-lqr-activation-steering-video-generation-reduced-order-linear-optimal-control] Local linearized dynamics within this subspace generalize across prompts, with Jacobian distances across 20 prompts far below the distance to random matrices [hong-etal-2026-la-lqr-activation-steering-video-generation-reduced-order-linear-optimal-control] An LQR feedback gain computes a closed-loop correction tracking the projected contrastive coordinate toward a setpoint, applying perturbation proportional to online tracking error [hong-etal-2026-la-lqr-activation-steering-video-generation-reduced-order-linear-optimal-control] Applied to Wan2.1-T2V-14B and HunyuanVideo, it reduces unsafe generations on T2VSafetyBench and SafeSora while preserving fidelity better than fixed-magnitude steering [hong-etal-2026-la-lqr-activation-steering-video-generation-reduced-order-linear-optimal-control] The experiment/model section was not fully reachable at render, so the exact model identities are deferred for re-verification [hong-etal-2026-la-lqr-activation-steering-video-generation-reduced-order-linear-optimal-control]

Context

contrastive-prompt-pair SVD subspace extraction, quantified via an energy-capture ratio, validated local-linear dynamics (Jacobian similarity across prompts) within the extracted subspace, closed-loop optimal-control (LQR) tracking of a subspace coordinate as a geometry-tied causal intervention

Papers

Activation Steering of Video Generation Models via Reduced-Order Linear Optimal Control — Hong, Jihoon, Chan, Alice, Dai, Qiyue, Skifstad, Julian, Chou, Glen2026 · arXiv:2606.04775