As noted in an answer to a previous question: we were originally targeting 24hr-turn games, but ended up pivoting to 5min-turn games due to the inability to gather a sufficient number of samples in the 24hr-turn format (as playing a single game can sometimes take months)! Playing 24hr-turn games would indeed pose additional challenges from a language generation perspective — while human players tend to send a similar number of messages in each format, messages in 24hr turns tend to be significantly longer (and likely more complex). Moreover, human players would have more time to interrogate mistakes from the bot, which could potentially lead to the agent making further mistakes. -ED