MPR-CiteG: Enhancing RAG with Multi-Portfolio Retrieval and Citation-Grounded Generation
The paper introduces MPR-CiteG, a dual-component framework that combines Multi-Portfolio Retrieval and Citation-Grounded Generation to enhance RAG systems by improving retrieval efficiency and ensuring factual consistency through explicit source attribution, thereby reducing hallucinations and achieving second place in the ScienceON AI Challenge.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to write a school report on a super complicated topic, like how to build a robot that can dance. You have a massive library of books, but they are all mixed up, and some of them are written in different languages. If you just ask a super-smart AI to write the report from its own memory, it might make up cool-sounding facts that sound true but are actually wrong. This is called "hallucination," and it's a big problem when you need real, trustworthy answers. To fix this, scientists use a trick called Retrieval-Augmented Generation (RAG). Think of RAG as giving the AI a "cheat sheet" of real documents to read before it writes. But here's the catch: if the AI picks the wrong cheat sheets, or if it writes the report without telling you exactly which page the information came from, the answer is still useless. The goal is to build an AI that is not only smart but also a careful researcher who always cites its sources.
This is exactly what the MPR-CiteG framework does. The researchers behind this project, who came in second place in a major AI challenge, built a two-part system to make sure AI answers are both accurate and honest. First, they created a Multi-Portfolio Retriever (MPR). Instead of asking the library for help with just one simple question, the MPR acts like a team of four different detectives, each with a unique strategy. One detective expands the question to find related ideas, another looks for synonyms to catch different ways of saying the same thing, a third balances the search between broad and specific, and the fourth translates the question into other languages to find international research. They all dig for clues, combine their findings, and then a strict editor (a re-ranker) picks the very best 50 documents to use.
Once the best documents are gathered, the second part of the system, called Citation-Grounded Generation (CiteG), takes over. This is the writer who refuses to guess. It reads the top 10 documents and writes a clear, fluent answer. But here is the magic: for every single sentence that makes a specific fact, it immediately attaches a tiny tag pointing to the exact source document. If a sentence is just a general transition, it doesn't get a tag. This way, you can read the answer and instantly know, "Okay, this fact comes from Source A, and this one from Source B." The team tested this on a dataset of scientific questions and found that their system was highly effective at avoiding made-up facts. They discovered that using a specific type of AI model (Qwen-2.5-14B) worked best, and that having all four "detective" strategies in the retrieval team was crucial for getting the right information. While the system is very good, the authors admit it can sometimes struggle with very complex questions that need a back-and-forth thinking process, suggesting that future versions might need to be even more dynamic. Ultimately, MPR-CiteG shows us how to build AI that doesn't just talk, but actually knows what it's talking about and can prove it.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.