README.md

---
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**](https://huggingface.co/papers/2509.07485)
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
- [`MVP-base`](https://huggingface.co/Jun421/MVP-base)
- [`MVP-3b`](https://huggingface.co/Jun421/MVP-3b)

## Datasets

### Evaluation Datasets
- [BM25-Top100](https://huggingface.co/datasets/Soyoung97/beir-eval-bm25-top100) (`Soyoung97/beir-eval-bm25-top100`)

### Training Datasets
- [Train/Valid](https://huggingface.co/datasets/Jun421/MVP-train) (`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](https://github.com/soyoung97/ListT5) repository.