IterCOMP Framework for Adaptive Prompt Compression in Multi-hop QA
August 17, 2026
IterCOMP is a training-free framework that uses an iterative compression loop to improve multi-hop question answering. By decomposing documents into evidence segments and generating targeted follow-up questions, it improves Exact Match and F1 scores on MusiQue, 2WikiMultiHopQA, and HotpotQA while reducing context length.
HOW THIS AFFECTS YOU
●
builderYou can use this method to reduce RAG latency and costs without sacrificing reasoning accuracy in complex queries.
●
researcherYou can explore this iterative approach to compression as an alternative to static single-turn methods.