[taylorius]

I could be wrong, but I think the only probabilistic component of an LLM is the statistical word fragment selection at the end. Assuming this is true, one could theoretically run the program multiple times, making different fragment choices. This (while horribly inefficient) would allow a sort of backtracking.