Differences in how a model behaves during training versus deployment, causing instability in reinforcement learning.