In 2023, Nelson Liu and others from Stanford University argued in the paper "Lost in the Middle" that model performance is significantly lower when relevant information is located in the middle of a long context compared to when it is at the beginning or end, forming a U-shaped performance curve.