The context windows of current mainstream large language models have expanded to hundreds of thousands or even millions of tokens, but under ultra-long contexts the model's attention weights on earlier information decay significantly (the "lost in the middle" problem).