Causal mask
Also: causal masking, autoregressive mask
A constraint that stops each position attending to anything after it, applied by forcing those scores to negative infinity before the softmax. It is what lets a model generate: during training, position t only ever sees what preceded it.
What people get wrong
Break it and training loss falls suspiciously fast, because the model is reading the answer. A loss far below what the task should allow is nearly always a broken mask.
A definition is not an understanding. This is taught properly in Attention, conceptually, part of Transformers & attention.
Learn this properly