This page is about a project by Jiayuan Liu and Vincent Conitzer that allows anyone to generate text by having multiple models take turns in generating tokens. The main purpose of this page is to give the link to the current location of this demo; currently, you can find it here.

Specifically, this demo allows multiple LLMs to alternatingly generate response tokens based on an initial user prompt. After selecting the models, setting the number of tokens per turn, and choosing the maximum number of turns, the user starts the process. The selected models then take turns extending a shared piece of text. Internally, the system maintains a single evolving document that all models can see. On each turn, the backend calls one model with the full text so far, along with an instruction such as: "Your task is to continue the following piece of writing... Do not repeat any existing text---only add new content to continue. {followed by the text generated so far}" The model's continuation is appended to the shared context, and control then rotates to the next model.