Cerebellar interneurons execute a low-dimensional control policy for real-time kinematic calibration

Abstract Artificial agents for continuous control currently struggle with the computational and latency overhead of processing high-dimensional sensory data. How biological neural networks perform similar feedback-control tasks with ultra-efficiency remains unclear. Here, we reveal a neurocomputational architecture in the mammalian cerebellum that resolves this bottleneck. Using in vivo two-photon calcium imaging in awake-behaving mice, we demonstrate that the cerebellar molecular layer interneuron (MLI) network acts as a native dimension-reduction layer. The MLI population distills high-dimensional somatosensory feedback into a continuous, linear scalar variable for locomotor speed. Operating without reliance on algorithmic trial-averaging, this network achieves 95% single-trial inference accuracy with 50-ms latency, yielding a strict mathematical control policy: zMLI(t)=v(t−τ). Bidirectional optogenetic modulations demonstrate that this interpretable biological latent space provides immediate, causal feedback calibration of motor output. These findings unearth a fundamental biological algorithm for real-time kinematic control, providing a deterministic blueprint for physics-informed neural networks and biomimetic robotic architectures.