中间迷失问题及为何检索优于填充

Research indicates that language models face the issue of ignoring information placed in the middle of long contexts. In 2023, researchers from Stanford, UC Berkeley, and Samaya AI published a paper titled 'Lost in the Middle: How Language Models Use Long Contexts,' which tested model performance based on the position of key information in the input. The findings revealed that models performed best when crucial information was located at the beginning or the end of the context, while information in the middle was often disregarded. This was not an isolated finding; the research team conducted tests across various model families and context lengths, consistently highlighting the significance of information positioning when utilizing long contexts.