The example in the "Demo IV" video seems to be like that. I don't know Chinese, but this is directly from the video: – Word-by-Word Translation: river zemin

The example in the "Demo IV" video seems to be like that. I don't know Chinese, but this is directly from the video: – Word-by-Word Translation: river zemin correct united states president of speak express regret Simultaneous Translation (wait 3): jiang zemin expressed his welcome to the us president 's remarks . Simultaneous Translation (wait 5): jiang zemin expressed his regret over the us president 's remarks . Baseline Translation (greedy): jiang zemin expressed regret over the us president 's remarks . – The "wait 3" model and "wait 5" model have a fixed 3 and 5 words, respectively, of "buffer" before they have to start translating. In this case, the Chinese word corresponding to "regret" comes between 3 and 5 words after its corresponding position in the English sentence. "Wait 5" is able to see it, but "wait 3" simply has to guess what goes in that position – and what it guesses is almost the exact opposite of the actual meaning! And it doesn't seem to be able to revise its translation after the fact. To be useful in practice, you would definitely need to be able to do that. It seems like it would also help to dynamically vary the buffer size based on the amount of ambiguity – but according to the paper on arxiv, previous systems already did that, while theirs intentionally throws it away in favor of a fixed buffer, apparently because it's simpler to train and because they prioritize low latency.