[bobbiechen]

> When people play Avalon, the games are usually rich in “cheap talk,” such as defending oneself, accusing others, or debunking others’ claims [10]. In this work, we do not consider the strategic implications of natural language communication...

In the 2189 mixed human/agent games we collected, all humans knew which players were human and which were DeepRole. There were no restrictions on chat usage for the human players, but DeepRole did not say anything and did not process sent messages

I think the fun part about Avalon and other hidden role games really comes from the "cheap talk", where people try to convince each other that their picks / approvals make sense as a member of the good team, as opposed to making the decisions from picks and approvals alone. Though from the results it seems that those concrete actions are already enough to outperform the humans.

There's also the consideration that because the humans know the identity of DeepRole as a bot, they play differently: "That's what I'm gonna pick, because that's what bots do" [1]. I wonder if a combined DeepRole + human-for-chatting-only team would outperform either alone.

[1] From Appendix F of the paper: https://www.youtube.com/watch?v=9RkUFHYTo_s