# EfficientRAG: Efficient Retriever for Multi-Hop Question Answering

**Authors:** Ziyuan Zhuang, Zhiyang Zhang, Sitao Cheng, Fangkai Yang, Jia Liu, Shujian Huang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang, Qi Zhang  
**arXiv:** [2408.04259](https://arxiv.org/abs/2408.04259)  
**Date:** 2024-08-08  
**Summary page:** https://www.summarizepaper.com/en/paper/2408.04259v1/

## Key points

- EfficientRAG is a retriever designed for multi-hop question answering, addressing challenges faced by retrieval-augmented generation (RAG) methods.
- It offers iterative query generation without repeated calls to large language models (LLMs), filtering out irrelevant information and refining queries for enhanced performance.
- Unlike other iterative retrieval methods, EfficientRAG streamlines the process without frequent LLM calls, reducing latency and cost.
- Experimental results show that EfficientRAG outperforms existing RAG methods on three open-domain multi-hop question-answering datasets.
- The research team behind EfficientRAG includes Ziyuan Zhuang, Zhiyang Zhang, Sitao Cheng, Fangkai Yang, Jia Liu, Shujian Huang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang,and Qi Zhang.

## Summary

, , , , 

EfficientRAG is a new and improved retriever designed specifically for multi-hop question answering. It addresses the challenges faced by retrieval-augmented generation (RAG) methods when dealing with complex queries, such as those involving multiple steps or hops. Unlike traditional one-round RAG methods that may struggle with these types of questions, EfficientRAG offers a solution through iterative query generation without the need for repeated calls to large language models (LLMs). This approach filters out irrelevant information and refines queries in an iterative manner, resulting in enhanced performance on multi-hop question answering tasks. The field of multi-hop question answering has seen various approaches, including multiple rounds of retrievals or reasonings, query rewriting or generation for subsequent rounds of retrieval, interleaving retrieval and reasoning steps, and self-asking mechanisms. However, these iterative retrieval methods often rely on multiple LLM calls for query manipulation and may require specific prompts and few-shot examples, leading to increased latency and cost. In contrast, EfficientRAG streamlines the process of gathering relevant information for multi-hop question answering without the overhead of frequent LLM calls. Experimental results have shown that EfficientRAG outperforms existing RAG methods on three open-domain multi-hop question-answering datasets. The research team behind this innovative approach includes Ziyuan Zhuang, Zhiyang Zhang, Sitao Cheng, Fangkai Yang, Jia Liu, Shujian Huang, Qingwei Lin, Saravan Rajmohan, Dongmei Zhang,and Qi Zhang. Their work represents a significant step towards more efficient and effective approaches to handling complex queries in natural language processing tasks.