持续预训练/RL降低大模型事实性，仅Muse Spark例外

Horrifying. Do you see? Except for Muse Spark, continued pretraining/RL degrade factuality in major model series. Gemini, GPT, Opus. But iirc this is not true for GLM-5, Minimax, Xiaomi. There clearly are different regimes.