BRIEF
# BRIEF-PRO: Universal Context Compression with Short-to-Long Synthesis for Fast and Accurate Multi-Hop Reasoning
[Jia-Chen Gu*](https://jasonforjoy.github.io/), [Junyi Zhang*](https://levi-zjy.github.io/), [Di Wu](https://xiaowu0162.github.io/), [Yuankai Li](https://lieziwind.github.io/), [Kai-Wei Chang](http://web.cs.ucla.edu/~kwchang/) and [Nanyun Peng](https://violetpeng.github.io/)
*Equal contribution
[**🌐 Homepage**]() | [**🤗 Model**](https://huggingface.co/uclanlp/brief-pro) | [**🤗 Dataset**](https://huggingface.co/datasets/uclanlp/Brief-Pro) | [**📖 Paper**]()
News
Intro
- BRIEF-PRO. It is a universal, lightweight compressor that distills relevant evidence for a given query from multiple retrieved documents into a concise summary for seamless integration into in-context RAG.
- Given seed data with contexts that are relatively short, each containing fewer than 1k words, BRIEF-PRO is trained towards abstractively condensing long contexts of 10k+ words across diverse scenarios.
- Furthermore, BRIEF-PRO offers flexible user control over the length of the output summary by allowing users to specify the desired number of sentences.
</a>
- The overview of the synthetic data pipeline for training BRIEF-PRO is as follows. Starting with a mixture of oracle and distractor documents for a given query, the pipeline: (a) expands each document by looking up an external knowledge corpus, (b) curates a compact summary by further reducing redundancy in the oracle documents, and (c) generates a user-controllable compression instruction by counting the number of sentences in the compact summary.
</a>
Training
- Coming Soon
Evaluation
- Coming Soon