优化训练策略的幻象：单调推理策略才是LLM强化学习的真正目标

The Mirage of Optimizing Training Policies

Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning