This isn't really true. The algorithm that generates your output, of which the LLM is a piece, is often beam search. Beam search can "back up" because it creates a branching tree of possible outputs and prunes them as it finds better ones
Similar idea pruning happens in reasoning models
