paper / jimmybajimmyba / May 12 / failedEMA Policy Gradient: Taming Reinforcement Learning for LLMs with EMA Anchor and Top-k KL