Transformers are Inherently Succinct
LLMs Get Lost in Multi-Turn Conversation：为什么大模型在多轮对话里会越聊越偏？