昨天，斯坦福大学、加州伯克利大学和Samaya AI的研究人员联合发布的一个论文中有一个非常有意思的发现:当相关信息出现在输入上下文的开始或结束时，大模型的性能通常最高，而当大模型必须访问长上下文中间的相关信息时，性能显著下降。
两种数据集的结果一致，下面以 BioS 为例，展示一个样例条目:Anya Briar Forger was born on October 2， 1996. She spent her early years in Princeton， NJ. She received mentorship and guidance from faculty members at MIT. She completed her education with a focus on Communications. She had a professional role at Meta Platforms. She was employed in Menlo Park， CA.
大语言模型的「推理」能力应该不是推理，在今年 6 月，一篇 Nature 论文《Language is primarily a tool for communication rather than thought》曾引发 AI 社区的大讨论，改变了我们对于 AI 智力的看法。
