IterCOMP: Training-Free Adaptive Prompt Compression for Multi-Hop QA
A novel framework named IterCOMP has been unveiled to tackle the difficulties associated with multi-hop question answering in retrieval-augmented generation systems. This approach, outlined in a paper available on arXiv (2608.13588), introduces a unified, training-free method for prompt compression that incorporates multi-hop reasoning within an iterative process. IterCOMP breaks down documents into segments of evidence, assesses the answerability of questions, and formulates specific follow-up queries to systematically gather crucial evidence, thereby creating a concise, reasoning-focused prompt. Testing on MusiQue, 2WikiMultiHopQA, and HotpotQA indicates significant enhancements in Exact Match and F1 scores while minimizing context length. This research is pertinent to AI and natural language processing, especially for boosting efficiency and precision in intricate question answering tasks.
Key facts
- IterCOMP is a training-free prompt compression framework.
- It incorporates multi-hop reasoning within an iterative compression loop.
- It decomposes documents into evidence segments.
- It evaluates question answerability and generates follow-up questions.
- Experiments were conducted on MusiQue, 2WikiMultiHopQA, and HotpotQA.
- IterCOMP improves Exact Match and F1 scores.
- It reduces context length.
- The paper is available on arXiv with ID 2608.13588.
Entities
Institutions
- arXiv