Research map / Large language models / Retrieval and knowledge
RetroLM: Retrieval-Augmented KVs for Long-Context Processing
Kun Luo, Zheng Liu, Shitao Xiao, Jiabei Chen, Hongjin Qian, Peitian Zhang, Shanshan Jiang, Bin Dong, Jun Zhao, Kang Liu
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 trained page retriever (Stage-1) (ablation) | RetroLM · LongBench QA (6 tasks) | average score: 29.6 -> 36.8 | Improved |
| add Stage-2 post-training on unsupervised text (ablation) | RetroLM-Stage1 (Mistral) · LongBench (12 tasks) | average score: 41.9 -> 43.5 | Improved |
| add Stage-2 post-training on unsupervised text (ablation) | RetroLM-Stage1 (Llama-3) · LongBench (12 tasks) | average score: 42.3 -> 44.4 | Improved |
| replace text-level RAG (BGE) with KV page retrieval | Mistral-7B · LongBench QA (6 tasks) | average score: 31.4 -> 36.8 | Improved |
| replace heuristic SnapKV with trained page retriever | Mistral-7B · LongBench (12 tasks) | average score: 39.4 -> 41.9 | Improved |
| full-attention fine-tuning on same data instead of Stage-2 (ablation) | Mistral-7B · LongBench QA (6 tasks) | average score: 33.1 -> 33.9 | No clear effect |
| InfLLM on fine-tuned model instead of RetroLM (ablation) | Mistral-7B · LongBench QA (6 tasks) | average score: 32.5 vs 39.1 | Hurt |
| increase KV budget 512 -> 1024 -> 2048 (ablation) | RetroLM-Stage1 · LongBench QA (6 tasks) | average score: 32.3 -> 34.2 -> 36.8 | Improved |
| RetroLM-Stage2 vs SnapKV on ultra-long QA | Mistral-7B · InfiniteBench QA | score: 16.2 -> 20.2 | Improved |
| RetroLM-Stage2 vs SnapKV on ultra-long summarization | Mistral-7B · InfiniteBench Summary | score: 25.3 -> 29.2 | Improved |
| add Stage-2 training | Mistral-7B-Instruct · MMLU | accuracy: 62.5 -> 63.2 | No clear effect |
| add Stage-2 training | Mistral-7B-Instruct · GSM8K | accuracy: 35.4 -> 34.6 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- InfLLM — InfLLM: Training-Free Long-Context Extrapolation for LLMs with an Efficient Context Memory (NeurIPS 2024) · better on 3 of 3 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.92)
- RAPID: Long-Context Inference with Retrieval-Augmented Speculative Decoding (ICML 2025 · similarity 0.92)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.91)
- Retrieval meets Long Context Large Language Models (ICLR 2024 · similarity 0.91)
- Benchmarking Large Language Models in Retrieval-Augmented Generation (AAAI 2024 · similarity 0.90)
- Accelerating Inference of Retrieval-Augmented Generation via Sparse Context Selection (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.
