Giving AI space to represent and deliberate about the world separately (secretly, in the case of Diplomacy) is an obvious, but very productive step once “think

Giving AI space to represent and deliberate about the world separately (secretly, in the case of Diplomacy) is an obvious, but very productive step once “think step by step” has been established as a key improvement over LLM’s standard logorrhea (not dissing, it’s just their only way of interacting with the world: spewing the next word, again and again). I’m curious if there’s more to this model than a turn management, input and output world states and prompting two streams of consciousness and using one to inform the other at every step. More new models have required some creative tricks to not be disappointing in “obvious” (human common sense) ways.