Countering Reward Over-Optimization in LLM with Demonstration-Guided Reinforcement Learning F i n d i n g s - A C L