TLCR: Token-Level Continuous Reward for Fine-grained Reinforcement Learning from Human Feedback F i n d i n g s - A C L