#exposure-bias

Question 1: Single choice

During seq2seq training you replace teacher forcing by feeding the decoder its own prediction from the previous step. What is the main trade-off?

Training matches test-time use, but the decoder first learns from its own errors, so learning is slower1
Training gets faster, because the decoder no longer needs the target sentence2
Nothing changes, because the decoder sees exactly the same tokens whenever its predictions are correct3
The decoder can now look at future target tokens, which lowers the training loss4