# BRIEF-Pro: Universal Context Compression with Short-to-Long Synthesis for Fast and Accurate Multi-Hop Reasoning

**Authors:** Jia-Chen Gu, Junyi Zhang, Di Wu, Yuankai Li, Kai-Wei Chang, Nanyun Peng  
**arXiv:** [2510.13799](https://arxiv.org/abs/2510.13799)  
**Date:** 2025-10-15  
**Summary page:** https://www.summarizepaper.com/en/paper/2510.13799v1/

## Key points

- Authors introduce BRIEF-Pro as a solution for challenges in retrieval-augmented generation (RAG) for multi-hop reasoning tasks
- BRIEF-Pro is a universal and lightweight compressor that distills relevant evidence from retrieved documents into concise summaries
- Model trained using short contexts to compress extended contexts exceeding 10k words
- Users can control the length of the summary by specifying the desired number of sentences
- Experiments show BRIEF-Pro generates more concise and relevant summaries compared to existing methods, enhancing performance across different language models

## Summary

, , , , 

In their paper titled "BRIEF-Pro: Universal Context Compression with Short-to-Long Synthesis for Fast and Accurate Multi-Hop Reasoning," authors Jia-Chen Gu, Junyi Zhang, Di Wu, Yuankai Li, Kai-Wei Chang, and Nanyun Peng introduce a novel approach to address the challenges posed by retrieval-augmented generation (RAG) in handling complex tasks. They highlight that while expanded contexts provide valuable information, they also come with drawbacks such as increased latency and cognitive load on the model. To overcome these limitations, especially in the context of intricate multi-hop questions, the authors propose BRIEF-Pro. <BRIEF-Pro></BRIEF-Pro> is described as a universal and lightweight compressor designed to distill relevant evidence from retrieved documents into concise summaries that can seamlessly integrate into in-context RAG. The model is trained using seed data with short contexts (less than 1k words) to perform abstractive compression of extended contexts exceeding 10k words across various scenarios. Notably, BRIEF-Pro offers users flexibility in controlling the length of the summary by allowing them to specify the desired number of sentences. The authors conducted experiments on four open-domain multi-hop question-answering datasets to evaluate the performance of BRIEF-Pro. The results show that BRIEF-Pro generates more concise and relevant summaries compared to existing methods, thereby enhancing performance across different language models – including small, large, and proprietary ones. In particular, when tested with the 70B reader model, BRIEF-Pro achieved a significant improvement in QA performance by achieving 32x compression over LongLLMLingua's 9x compression while requiring only 23% of its computational overhead. This demonstrates the effectiveness of BRIEF-Pro in efficiently summarizing complex information for enhanced multi-hop reasoning tasks.