Research map / Large language models / Retrieval and knowledge
GRAPE: Let GRPO Supervise Query Rewriting by Ranking for Retrieval
Zhaohua Zhang, Jianhuan Zhuo, Muxi Chen, Chenchen Zhao, Wenyu Jiang, Mingyang Chen, Yu Tang, Qiuyong Xiao, Tianwen Jiang, Jihong Zhang, Zhixun Su
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune the query rewriter with ranking-reward GRPO instead of a frozen LLM rewriter | CLIP ViT-B/32 · Wikipedia | R@10 (%): 67.1 -> 78.4 | Improved |
| fine-tune the query rewriter with ranking-reward GRPO instead of a frozen LLM rewriter | CLIP ViT-L/14 · Flickr30k-CN | R@1 (%): 57.6 -> 62.6 | Improved |
| fine-tune the query rewriter with ranking-reward GRPO instead of a frozen LLM rewriter | CLIP ViT-B/32 · CVLUE | R@1 (%): 10.6 -> 13.1 | Improved |
| use GRAPE rewriting across all three CLIP backbones | CLIP + Qwen2.5-3B-Instruct rewriter · average of five benchmarks | R@10 (%): 67.7 -> 74.2 (ViT-B/32) | Improved |
| replace ranking-based reward with similarity-score reward (ablation) | GRAPE · Flickr30k-CN | R@1 | Hurt |
| use corpus-relative ranking reward instead of similarity reward (ablation) | GRAPE · Flickr30k-CN | R@1 | Improved |
| train on 10k samples instead of 100k (ablation) | GRAPE · Flickr30k-CN | R@1: 94% of full-data R@1 | No clear effect |
| train on 10k samples instead of full data (ablation) | GRAPE · CVLUE | R@1: 88% of full-data performance | Mixed |
Closest papers by meaning
- Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.87)
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.87)
- MaFeRw: Query Rewriting with Multi-Aspect Feedbacks for Retrieval-Augmented Large Language Models (AAAI 2025 · similarity 0.87)
- Contrastive Retrieval Heads for Improved Attention-Based Reranking (NeurIPS 2026 · similarity 0.86)
- AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking (NeurIPS 2025 · similarity 0.86)
- RankRAG: Unifying Context Ranking with Retrieval-Augmented Generation in LLMs (NeurIPS 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
