CtrlCache Paper Cuts Interactive Video World Model Compute 1.21–1.41x Using Control-Aware Caching

CtrlCache is a training-free caching scheme for interactive video world models that uses pre-arriving user control signals to label each chunk as initial, transition, turning or steady, then reuses transformer residuals in low-change regimes. On Matrix-Game 2.0 and LingBot-World v1/v2, it delivers 1.21x–1.41x speedups by exploiting that low-frequency structure persists while high-frequency detail changes around action transitions.