GroupRank: A Groupwise Paradigm for Effective and Efficient Passage Reranking with LLMs

Researchers introduce GroupRank, a novel LLM-based passage reranking paradigm that balances efficiency and accuracy by combining pointwise and listwise ranking approaches. The method achieves state-of-the-art performance with 65.2 NDCG@10 on BRIGHT benchmark while delivering 6.4x faster inference than existing approaches.

GroupRank addresses a fundamental limitation in how large language models handle information retrieval tasks. Traditional reranking approaches force researchers to choose between efficiency and accuracy: pointwise methods quickly score individual documents but miss comparative context, while listwise methods capture full ranking context but consume excessive computational resources and tokens. This research introduces a middle path through groupwise ranking, where documents are processed in strategic batches to maintain contextual awareness without the prohibitive overhead of full listwise evaluation.

The innovation extends beyond architecture alone. The researchers developed an answer-free data synthesis pipeline that combines local pointwise signals with global ranking patterns, enabling effective model training through both supervised fine-tuning and reinforcement learning. A specialized reward function balances ranking utility and group alignment, optimizing both document ordering and score calibration simultaneously. This dual-optimization approach addresses why previous methods struggled: they either ignored relevance calibration or failed to properly order documents relative to peers.

The practical impact targets both academic and commercial information retrieval systems. The 6.4x speedup makes LLM-based reranking economically viable for production search infrastructure, where inference costs directly impact operational budgets. Performance gains of 2.1 NDCG@10 points on medical literature datasets (R2MED) demonstrate improved handling of complex domain-specific queries, benefiting specialized search applications in healthcare, legal, and scientific research. These improvements accumulate across millions of queries, creating measurable user experience enhancements.

Future developments will likely explore whether groupwise paradigms extend effectively to other ranking problems and whether the data synthesis approach generalizes across different LLM architectures. The research suggests opportunities for hybrid ranking systems that adaptively select between pointwise, groupwise, and listwise approaches based on query complexity and computational constraints.

• →GroupRank achieves 65.2 NDCG@10 on BRIGHT benchmark, setting new state-of-the-art for LLM-based passage reranking
• →The groupwise approach delivers 6.4x inference speedup compared to listwise methods while maintaining superior ranking accuracy
• →Answer-free data synthesis pipeline enables effective training by fusing pointwise and listwise ranking signals
• →Specialized group-ranking reward function optimizes both document ordering and relevance score calibration simultaneously
• →Framework demonstrates 2.1 point gains on domain-specific benchmarks like R2MED, suggesting practical value for specialized search applications