| --- |
| 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](https://huggingface.co/papers/2508.16757). |
|
|
| Code: [https://github.com/DataScienceUIBK/llm-reranking-generalization-study](https://github.com/DataScienceUIBK/llm-reranking-generalization-study) |
|
|
| Project Page / Leaderboard: [https://rankarena.ngrok.io](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](https://arxiv.org/abs/2508.05512) leaderboard. You can view and interact with RankArena through this [link](https://rankarena.ngrok.io) |
| - **[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. |
|
|
| <div align="center"> |
| <img src="https://github.com/DataScienceUIBK/llm-reranking-generalization-study/blob/main/figures/radar.jpg" alt="Performance Overview" width="600"/> |
| <p><em>Performance comparison across pointwise, pairwise, and listwise reranking paradigms</em></p> |
| </div> |
|
|
| ### 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) |
|
|
| <div align="center"> |
| <img src="https://github.com/DataScienceUIBK/llm-reranking-generalization-study/blob/main/figures/efficiency_tradeoff.png.jpg" alt="Efficiency Analysis" width="700"/> |
| <p><em>Runtime vs. Performance trade-offs across reranking methods</em></p> |
| </div> |
|
|
| # 🔧 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](https://github.com/DataScienceUIBK/llm-reranking-generalization-study/blob/main/SUBMISSION.md) for guidelines. |
|
|
| Current standings available at: [RanArena](https://rankarena.ngrok.io) |
|
|
| # 🤝 Contributing |
|
|
| We welcome contributions! See [CONTRIBUTING.md](https://github.com/DataScienceUIBK/llm-reranking-generalization-study/blob/main/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: |
|
|
| ```bibtex |
| @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](mailto:abdelrahman.abdallah@uibk.ac.at), [Bhawna Piryani](mailto:bhawna.piryani@uibk.ac.at) |
| - **Institution**: University of Innsbruck |
| - **Issues**: Please use GitHub Issues for bug reports and feature requests |
|
|
| --- |
|
|
| <div align="center"> |
| <p>⭐ Star this repo if you find it helpful! ⭐</p> |
| <p>📧 Questions? Open an issue or contact the authors</p> |
| </div> |
