Inverse-Q*: Token Level Reinforcement Learning for Aligning Large Language Models Without Preference Data F i n d i n g s - E M N L P