Paper 2511.11653
GroupRank: A Groupwise Paradigm for Effective and Efficient Passage Reranking with LLMs
- Published
- Nov 2025
- Research lab
- Independent
- Citations
- 3
- GitHub
- 259 stars
01 In brief
Summary
GroupRank is a novel groupwise reranking paradigm for LLM-based passage retrieval that balances the efficiency of pointwise methods with the accuracy of listwise approaches.
It partitions candidate documents into small groups for parallel cross-document comparison, avoiding context-window limits and reducing latency.
Training uses an answer-free data synthesis pipeline that fuses pointwise and listwise teacher signals, followed by SFT and RL with a Group-Ranking reward combining ranking-utility (NDCG@10, Recall@10) and group-alignment (JS divergence) components.
Experiments show GroupRank achieves state-of-the-art NDCG@10 of 65.2 on BRIGHT and 52.3 on R2MED, outperforming baselines by up to 2.1 points, while delivering 2.4-6.4× faster inference.
Ablations confirm each component is essential, with SFT providing the foundation and RL boosting performance.
Random grouping outperforms sorted grouping, and test-time scaling further improves results.
02 From the paper
Abstract
Large Language Models (LLMs) have emerged as powerful tools for passage reranking in information retrieval, leveraging their superior reasoning capabilities to address the limitations of conventional models on complex queries. However, current LLM-based reranking paradigms are fundamentally constrained by an efficiency-accuracy trade-off: (1) pointwise methods are efficient but ignore inter-document comparison, yielding suboptimal accuracy; (2) listwise methods capture global context but suffer from context-window constraints and prohibitive inference latency. To address these issues, we propose GroupRank, a novel paradigm that balances flexibility and context awareness. To unlock the full potential of groupwise reranking, we propose an answer-free data synthesis pipeline that fuses local pointwise signals with global listwise rankings. These samples facilitate supervised fine-tuning and reinforcement learning, with the latter guided by a specialized group-ranking reward comprising ranking-utility and group-alignment. These complementary components synergistically optimize document ordering and score calibration to reflect intrinsic query-document relevance. Experimental results show GroupRank achieves a state-of-the-art 65.2 NDCG@10 on BRIGHT and surpasses baselines by 2.1 points on R2MED, while delivering a 6.4$\times$ inference speedup.