README.md

metadata

license: mit
task_categories:
  - text-ranking
language:
  - en
tags:
  - information-retrieval
  - reranking
  - llm
  - benchmark
  - temporal
  - llm-reranking

How Good are LLM-based Rerankers? An Empirical Analysis of State-of-the-Art Reranking Models 🔍

This repository contains the FutureQueryEval Dataset presented in the paper How Good are LLM-based Rerankers? An Empirical Analysis of State-of-the-Art Reranking Models.

Code: https://github.com/DataScienceUIBK/llm-reranking-generalization-study

Project Page / Leaderboard: https://rankarena.ngrok.io

🎉 News

• [2025-08-22] 🎯 FutureQueryEval Dataset Released! - The first temporal IR benchmark with queries from April 2025+
• [2025-08-22] 🔧 Comprehensive evaluation framework released - 22 reranking methods, 40 variants tested
• [2025-08-22] 📊 Integrated with RankArena leaderboard. You can view and interact with RankArena through this link
• [2025-08-20] 📝 Paper accepted at EMNLP Findings 2025

📖 Introduction

We present the most comprehensive empirical study of reranking methods to date, systematically evaluating 22 state-of-the-art approaches across 40 variants. Our key contribution is FutureQueryEval - the first temporal benchmark designed to test reranker generalization on truly novel queries unseen during LLM pretraining.

Performance comparison across pointwise, pairwise, and listwise reranking paradigms

Key Findings 🔍

• Temporal Performance Gap: 5-15% performance drop on novel queries compared to standard benchmarks
• Listwise Superiority: Best generalization to unseen content (8% avg. degradation vs 12-15% for others)
• Efficiency Trade-offs: Comprehensive runtime analysis reveals optimal speed-accuracy combinations
• Domain Vulnerabilities: All methods struggle with argumentative and informal content

📄 FutureQueryEval Dataset

Overview

FutureQueryEval is a novel IR benchmark comprising 148 queries with 2,938 query-document pairs across 7 topical categories, designed to evaluate reranker performance on temporal novelty.

🎯 Why FutureQueryEval?

• Zero Contamination: All queries refer to events after April 2025
• Human Annotated: 4 expert annotators with quality control
• Diverse Domains: Technology, Sports, Politics, Science, Health, Business, Entertainment
• Real Events: Based on actual news and developments, not synthetic data

📊 Dataset Statistics

Metric	Value
Total Queries	148
Total Documents	2,787
Query-Document Pairs	2,938
Avg. Relevant Docs per Query	6.54
Languages	English
License	MIT

🌍 Category Distribution

• Technology: 25.0% (37 queries)
• Sports: 20.9% (31 queries)
• Science & Environment: 13.5% (20 queries)
• Business & Finance: 12.8% (19 queries)
• Health & Medicine: 10.8% (16 queries)
• World News & Politics: 9.5% (14 queries)
• Entertainment & Culture: 7.4% (11 queries)

📝 Example Queries

🌍 World News & Politics:
"What specific actions has Egypt taken to support injured Palestinians from Gaza, 
as highlighted during the visit of Presidents El-Sisi and Macron to Al-Arish General Hospital?"

⚽ Sports:
"Which teams qualified for the 2025 UEFA European Championship playoffs in June 2025?"

💻 Technology:
"What are the key features of Apple's new Vision Pro 2 announced at WWDC 2025?"

Data Collection Methodology

1. Source Selection: Major news outlets, official sites, sports organizations
2. Temporal Filtering: Events after April 2025 only
3. Query Creation: Manual generation by domain experts
4. Novelty Validation: Tested against GPT-4 knowledge cutoff
5. Quality Control: Multi-annotator review with senior oversight

📊 Evaluation Results

Top Performers on FutureQueryEval

Method Category	Best Model	NDCG@10	Runtime (s)
Listwise	Zephyr-7B	62.65	1,240
Pointwise	MonoT5-3B	60.75	486
Setwise	Flan-T5-XL	56.57	892
Pairwise	EchoRank-XL	54.97	2,158
Tournament	TourRank-GPT4o	62.02	3,420

Performance Insights

• 🏆 Best Overall: Zephyr-7B (62.65 NDCG@10)
• ⚡ Best Efficiency: FlashRank-MiniLM (55.43 NDCG@10, 195s)
• 🎯 Best Balance: MonoT5-3B (60.75 NDCG@10, 486s)

Runtime vs. Performance trade-offs across reranking methods

🔧 Supported Methods

We evaluate 22 reranking approaches across multiple paradigms:

Pointwise Methods

• MonoT5, RankT5, InRanker, TWOLAR
• FlashRank, Transformer Rankers
• UPR, MonoBERT, ColBERT

Listwise Methods

• RankGPT, ListT5, Zephyr, Vicuna
• LiT5-Distill, InContext Rerankers

Pairwise Methods

• PRP (Pairwise Ranking Prompting)
• EchoRank

Advanced Methods

• Setwise (Flan-T5 variants)
• TourRank (Tournament-based)
• RankLLaMA (Task-specific fine-tuned)

🔄 Dataset Updates

FutureQueryEval will be updated every 6 months with new queries about recent events to maintain temporal novelty. Subscribe to releases for notifications!

Upcoming Updates

• Version 1.1 (December 2025): +100 queries from July-September 2025 events
• Version 1.2 (June 2026): +100 queries from October 2025-March 2026 events

📋 Leaderboard

Submit your reranking method results to appear on our leaderboard! See SUBMISSION.md for guidelines.

Current standings available at: RanArena

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for:

• Adding new reranking methods
• Improving evaluation metrics
• Dataset quality improvements
• Bug fixes and optimizations

🎈 Citation

If you use FutureQueryEval or our evaluation framework, please cite:

@misc{abdallah2025howgoodarellmbasedrerankers,
      title={How Good are LLM-based Rerankers? An Empirical Analysis of State-of-the-Art Reranking Models},
      author={Abdelrahman Abdallah and Bhawna Piryani},
      year={2025},
      eprint={2508.16757},
      archivePrefix={arXiv},
      primaryClass={cs.IR}
}

📞 Contact

• Authors: Abdelrahman Abdallah, Bhawna Piryani
• Institution: University of Innsbruck
• Issues: Please use GitHub Issues for bug reports and feature requests

---

⭐ Star this repo if you find it helpful! ⭐

📧 Questions? Open an issue or contact the authors