I plan to read the paper later, but anyone have a TL;DR for how they connected a language model to the game state? That seems like the real advance here. Language models are so prone to making stuff up and spouting nonsense, and controlling them is really hard.