This is just a weird feeling that I've been coming closer to articulating lately, but I only think that you can get forward reasoning from what is basically word association; there's no mechanism for unwinding it because it has no real memory. By "it" I mean word association itself, not any context window. It predicts what could be in a position, and ignores what wasn't in a position.

People don't do that. People are constantly engaging with paths not chosen. Right after I choose to write one thing, I'm immediately engaging with what I chose not to write there - I'm explaining why I didn't write it, I'm realizing that my choice may seem unusual so I'm trying to make it memorable, I'm focusing on the distinctions between what I wrote and what I didn't.

LLMs don't currently do that. LLMs just ape a structure. When the structure resembles the sort of timid, clarifying fussing I just described, the LLMs just drift randomly because what they didn't say wasn't in the context.

I also think that's why they have such a serious problem backtracking. They're not taking into account the already eliminated possibilities. Often the thing that was so unlikely that you weren't going to waste time on it is the answer, and things you discover while going down an ultimately wrong (but initially far more promising) path remind you of the path not taken.

They're simply assembling a thing that resembles a valid argument, and happen to make sound choices because the plurality of input happened to contain sound choices. This is usually a very good bet because there are so many more ways to be wrong than to be right. But it doesn't account for attractive (common) wrong choices. You need a way to back out of those.