[2404.19409] Countering Reward Over-optimization in LLM with Demonstration-Guided Reinforcement Learning