AssistanceZero is an algorithm published in 2025 by Stuart Russell's group at the University of California, Berkeley, for solving assistance games at scale, where only the human knows the true goal and the assistant must cooperate without knowing it. It combines AlphaZero-style search with a learned model of human behavior so that the assistant infers the goal from the human's actions and adjusts its own behavior accordingly.
The researchers evaluated the method in a Minecraft environment in which an assistant helps a human build a structure. Because the assistant learns to maintain uncertainty about the goal and adapt to human feedback, the approach differs from standard reinforcement learning, which optimizes a fixed reward.