How can machines figure out what goal an animal switches to mid-action?

Traditional inverse reinforcement learning assumes an animal pursues one fixed goal throughout an episode, but creatures (and robots) actually switch targets mid-task. PRISM uses a recurrent network to infer which goal an agent is pursuing at each moment from its observation history, then recovers the reward for each goal in closed form. Tested on mice in mazes and robots manipulating objects, PRISM recovered interpretable, time-coherent goals from unlabeled video—suggesting goal-switching is a real feature of both biological and artificial behavior.