Research map / Large language models / Retrieval and knowledge
Less Is More: Elevating RAG via Performance-Driven Context Compression
Ziqiang Cui, Yunpeng Weng, Xing Tang, Peiyang Liu, Shiwei Li, Bowei He, Jiamin Chen, Yansen Zhang, Xiuqiang He, Rui Zhang, Chen Ma
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 |
|---|---|---|---|
| add performance-driven RL compression instead of full top-5 documents | RAG with Qwen2.5-14B-Instruct · NQ | EM: 38.03 -> 41.02 | Improved |
| add performance-driven RL compression instead of full top-5 documents | RAG with Qwen2.5-14B-Instruct · 2WikiMultihopQA | EM: 29.64 -> 36.72 | Improved |
| add distillation warm-start before GRPO (ablation) | CORE · NQ | EM: 36.37 -> 41.02 | Improved |
| add RL stage after distillation (ablation) | CORE · NQ | EM: 34.18 -> 41.02 | Improved |
| add RL stage after distillation (ablation) | CORE · 2WikiMultihopQA | EM: 30.25 -> 36.72 | Improved |
| add distillation warm-start before GRPO (ablation) | CORE · TriviaQA | EM: 65.23 -> 65.63 | No clear effect |
| add F1 reward term (alpha > 0) to EM reward (ablation) | CORE reward · all four QA datasets | EM | Improved |
| apply compressor trained on top-5 docs to top-10 docs without retraining | CORE · NQ | EM: 38.67 -> 41.88 | Improved |
| transfer compressor to unseen Llama-3.1-8B-Instruct reader | CORE · NQ | EM: 37.65 -> 40.72 | Improved |
| train compressor on NQ and transfer to HotpotQA | CORE · HotpotQA | EM: 33.67 -> 32.15 | Mixed |
| increase compressor size from 1B to 3B (ablation) | CORE · QA benchmarks | EM | Improved |
| use off-the-shelf DeepSeek-V3 (671B) summaries instead of full documents | RAG with Qwen2.5-14B-Instruct · NQ | EM: 38.03 -> 37.73 | Hurt |
Closest papers by meaning
- Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection (ICLR 2025 · similarity 0.92)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.91)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.91)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.90)
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.90)
- RAG-DDR: Optimizing Retrieval-Augmented Generation Using Differentiable Data Rewards (ICLR 2025 · similarity 0.89)
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.
