According to StanfordAI Lab, Prefix Sliding reduces long-context costs and outperforms full attention and compaction on long reasoning tasks.