Tag
This paper proposes D-RPC, a method for distilling reasoning from large language models to smaller ones by compressing reasoning paths into a reusable bank, achieving better performance and consistency on math and commonsense benchmarks.