Looking for the earlier page? Examples from the first version of the paper.
Reviser is a language model that writes and edits text with cursor actions: at each step it inserts a token at the cursor, moves the cursor, deletes the token to the left of the cursor, or stops. This page replays real model outputs one action at a time, both for continuing a web-text prompt and for editing a short text by instruction. The examples were picked by hand, not sampled at random: for each pool, an automatic filter built 100 candidates (65 for the exact-match edit pools), an LLM reading agent picked 10, and the paper's Appendix J prints some of them (marked “Shown in Appendix J”). Sections 6 and 8 of the paper give the measured quality.
The prompts and edit items are public web and dataset text, and the outputs are model generations, so they can contain errors and made-up facts.
Each example opens on its final text. Use the buttons or the slider to replay it from the prompt or input draft, one action at a time. Text typed at the end of the text is shown plain; text typed with the cursor inside the text is highlighted.
The “Cursor moments” notes describe the key moves. Steps are numbered from 0. <-14 and ->2 are cursor jumps of 14 tokens left and 2 tokens right, | marks where the cursor lands in the quoted text, TYPE is what the model types there and DEL what it deletes; “backspace” is a delete. The step buttons under a replay show the text just before that step.
A move is a jump of 1, 2, 4, … tokens left or right; a longer jump is made of several moves. Tokens are GPT-2 tokens.