What does the causal attention mask in a decoder-only model prevent?
Answer: B. A token attending to tokens after it, which would leak the answer during training
Training predicts every next token in parallel. Without the mask, position t could see token t+1, the very thing it must predict.