Research map / Large language models / Retrieval and knowledge
Rethinking the Reranker: Boundary-Aware Evidence Selection for Robust Retrieval-Augmented Generation
Jiashuo Sun, Pengcheng Jiang, Saizhuo Wang, Jiajun Fan, Heng Wang, Siru Ouyang, Ming Zhong, Yizhu Jiao, Chengsong Huang, Xueqiang Xu, Pengrui Han, Peiran Li and 4 more
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 |
|---|---|---|---|
| remove boundary reward from selector (ablation) | BAR-RAG · NQ/PopQA/HotpotQA avg | exact match: 40.2 -> 44.2 | Improved |
| add relevance reward to selector (ablation) | BAR-RAG · NQ/PopQA/HotpotQA avg | exact match: 43.5 -> 44.2 | Improved |
| add citation reward to generator (ablation) | BAR-RAG · NQ/PopQA/HotpotQA avg | exact match: 42.7 -> 44.2 | Improved |
| add Stage 2 generator fine-tuning on selected evidence (ablation) | BAR-RAG · NQ/PopQA/HotpotQA avg | exact match: 36.8 -> 44.2 | Improved |
| add Stage 1 selector training (ablation) | BAR-RAG · NQ/PopQA/HotpotQA avg | exact match: 42.2 -> 44.2 | Improved |
| add Stage 1 selector training (ablation) | BAR-RAG · HotpotQA | exact match: 41.9 -> 38.8 | Hurt |
| add training-time filtering of trivial and unanswerable queries (ablation) | BAR-RAG selector · NQ/PopQA/HotpotQA avg | exact match: 35.6 -> 44.2 | Improved |
| increase iterations from 1 to 2 (ablation) | BAR-RAG · 7-benchmark avg, Qwen-2.5-7B | exact match: 37.2 -> 38.8 | Improved |
| increase iterations from 2 to 3 (ablation) | BAR-RAG · 7-benchmark avg, Qwen-2.5-3B | exact match: 34.3 -> 34.2 | No clear effect |
| add a neural reranker to standard RAG | RAG · 7-benchmark avg, Qwen-2.5-7B | exact match: 26.9 -> 28.0 | Improved |
| use BAR-RAG instead of RAG SFT | Qwen-2.5-7B-Instruct · 7-benchmark avg | exact match: 31.8 -> 39.1 | Improved |
| use IRCoT instead of standard RAG | Qwen-2.5-7B-Instruct · 7-benchmark avg | exact match: 26.9 -> 22.6 | Hurt |
Closest papers by meaning
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
- Retrieval is Not Enough: Enhancing RAG through Test-Time Critique and Optimization (NeurIPS 2025 · similarity 0.88)
- Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection (ICLR 2025 · similarity 0.88)
- Influence Guided Context Selection for Effective Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
- Worse than Zero-shot? A Fact-Checking Dataset for Evaluating the Robustness of RAG Against Misleading Retrievals (NeurIPS 2025 · similarity 0.88)
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.
