That's pretty interesting. Last year I messed with hooking language models up to inform 6/zmachines (both RL training and inference and even generating games) and have noticed most are awful at navigating and reasoning about graphs. Maybe this changed with the latest ones.