Research map / Large language models / Instruction and fine-tuning
ERank: Fusing Supervised Fine-Tuning and Reinforcement Learning for Effective and Efficient Text Reranking
Yuzheng Cai, Yanzhao Zhang, Dingkun Long, Mingxin Li, Pengjun Xie, Weiguo Zheng
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 |
|---|---|---|---|
| replace yes/no scoring with 0-10 integer scores (ablation) | QwQ-32B reranker · BRIGHT, TREC DL, BEIR-5 | nDCG@10: BRIGHT 20.8 -> 23.2; TREC DL 60.9 -> 66.1; BEIR-5 31.1 -> 37.1 | Improved |
| use 0-10 scores instead of 0-3 scores (ablation) | QwQ-32B reranker · BRIGHT, TREC DL, BEIR-5 | nDCG@10: 22.7 -> 23.2; 64.1 -> 66.1; 36.5 -> 37.1 | Improved |
| add RL stage after SFT (ablation) | ERank-4B · BRIGHT, BEIR-5, TREC DL | nDCG@10: BRIGHT 22.0 -> 22.7; BEIR-5 30.0 -> 32.4; TREC DL 68.1 -> 68.9; FollowIR 11.2 -> 11.0 | Mixed |
| add SFT before RL (ablation) | ERank-4B · BRIGHT, TREC DL | nDCG@10: RL only 20.1 -> 22.7 BRIGHT; 66.5 -> 68.9 TREC DL | Improved |
| use listwise-derived reward r_RR instead of pointwise squared-error reward r_SE (ablation) | ERank-4B RL · TREC DL | nDCG@10: 62.6 -> 68.9 | Improved |
| use r_RR instead of r_nDCG (ablation) | ERank-4B RL · BRIGHT, FollowIR | nDCG@10 / p-MRR: BRIGHT 21.5 -> 22.7; FollowIR 13.2 -> 11.0 | Mixed |
| add BM25 hybrid score aggregation | ERank-4B · BRIGHT (ReasonIR-8B retrieval) | nDCG@10: 30.5 -> 38.7 | Improved |
| scale ERank from 4B to 32B | ERank · BRIGHT | nDCG@10: 22.7 -> 24.4 (BM25 original queries) | Improved |
| use 3-sample consensus label selection vs single generation in data synthesis (ablation) | teacher labels (QwQ-32B) · 512-query sample | nDCG@10: 63.5 -> 65.6; 10 samples 67.4 | Improved |
Closest papers by meaning
- AcuRank: Uncertainty-Aware Adaptive Computation for Listwise Reranking (NeurIPS 2025 · similarity 0.89)
- Contrastive Retrieval Heads for Improved Attention-Based Reranking (NeurIPS 2026 · similarity 0.88)
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
- Attention in Large Language Models Yields Efficient Zero-Shot Re-Rankers (ICLR 2025 · similarity 0.88)
- Very Efficient Listwise Multimodal Reranking for Long Documents (ICML 2026 · similarity 0.87)
- TFRank: Think-Free Reasoning Enables Practical Pointwise LLM Ranking (AAAI 2026 · 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.
