The cerebellum implements structured representation of valence to support adaptive behavior control.

Animals are thought to form internal models in the brain representing safety and danger, enabling avoidance through learned associations. Although the cerebellum has been implicated in active avoidance, its role in constructing internal models remains unclear. Here, we combined calcium imaging and closed-loop virtual reality in adult zebrafish to examine cerebellar circuits underlying avoidance learning. Fish learned to associate visual cues with safety or danger in a cerebellum-dependent manner using electric shocks as instructive signals. Cerebellar output neurons and Purkinje cells in the dorsal cerebellum developed cue-selective responses. Output neurons encoded both valence states, with danger-preferring cells prevalent, whereas Purkinje cells showed stronger danger bias. After cue reversal, neural selectivity adapted to the new rule. Safety- and danger-selective output neurons were spatially segregated, indicating anatomically dedicated units for valence processing. Optogenetic inhibition of shock-related inputs to Purkinje cells impaired rule updating. Together, our results indicate that the cerebellum forms an internal model of safety and danger, enabling adaptive avoidance.