A training technique where parts of the input are hidden, and the model learns to predict what was masked, helping it understand underlying patterns.