- The problem
- A fast language-model shortcut guesses several next words at once, but guesses that each seem likely can still make an awkward or incorrect sequence together. Rechecking the sequence again and again can erase the speed benefit.
- The idea
- LiLiCorr keeps a shortlist of likely words for each position and scores how neighboring choices fit together, using information from the whole draft. For example, it can prefer “dog” followed by “barks” over individually likely words that do not form a sensible phrase. It scores all the options in one pass, then chooses the sequence from left to right.
- What it got
- Compared with the original shortcut, LiLiCorr increases the number of draft words accepted by 7–19%. In comparisons with other new methods, it is fastest in 63 of 72 tested settings.