I’m not familiar with the term “beam” in the context of LLMs, so that’s not factored into my argument in any way. LLMs generate text based on the history of tokens generated thus far, not just the last token. That is by definition non-Markovian. You can argue that an augmented state space would make it Markovian, but you can say that about any stochastic process. Once you start doing that, both become mathematically equivalent. Thinking about this a bit more, I don’t think it really makes sense to talk about a process being Markovian or not without a wider context, so I’ll let this one go.

nitpick that makes communication worse

How many readers do you think know what “Markov” means? How many would know what “stochastic” or “random” means? I’m willing to bet that the former is a strict subset of the latter.