Apple researchers target entropy collapse in reinforcement learning

Apple researchers propose entropy-preserving reinforcement learning methods to keep AI agents exploratory during training instead of collapsing too early into narrow behavior.