license: apache-2.0
library_name: transformers
pipeline_tag: text-ranking
MVP: Multi-view-guided Passage Reranking with Large Language Models
This repository contains the official implementation for the EMNLP 2025 paper: Multi-view-guided Passage Reranking with Large Language Models by Jeongwoo Na*, Jun Kwon*, Eunseong Choi, Jongwuk Lee (* : equal contribution)
Overview
MVP (Multi-View-guided Passage Reranking) is a non-generative LLM-based reranking method designed to overcome the efficiency and bias sensitivity challenges of existing LLM-based rerankers. It encodes query-passage information into diverse view embeddings, ensuring accurate representation without external biases. The model then combines query-aware passage embeddings to produce distinct anchor vectors, which are used to directly compute relevance scores in a single decoding step. An orthogonal loss encourages diversity across these views.
With just 220M parameters, MVP matches the performance of much larger 7B-scale fine-tuned models while achieving a 100x reduction in inference latency. The 3B-parameter variant of MVP achieves state-of-the-art performance on both in-domain and out-of-domain benchmarks.
How to Use
Setup Environment
conda env create -f mvp.yaml
conda activate mvp
Run MVP
cd inference
bash run_evaluation.sh
Model Checkpoints
Datasets
Evaluation Datasets
- BM25-Top100 (
Soyoung97/beir-eval-bm25-top100)
Training Datasets
- Train/Valid (
Jun421/MVP-train) This dataset is derived from the BEIR/MSMARCO license, and its usage is restricted to academic purposes only.
Acknowledgments
The implementation of this model is based on the ListT5 repository.