Token-Level Self-Play with Importance-Aware Guidance for Large Language Models | Read Paper on Bytez