Stabilizing Reinforcement Learning with LLMs: Formulation and Practices