Research map / Large language models / Retrieval and knowledge
Improving Long-Context Summarization with Multi-Granularity Retrieval Optimization
Xueyu Chen, Kaitao Song, Zifan Song, Dongsheng Li, Cairong Zhao
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 HTSIR summary tree on top of DPR retrieval (ablation) | DPR retriever + GPT-4o mini · QuALITY-HARD | accuracy: 55.31 -> 62.57 | Improved |
| add HTSIR summary tree on top of BM25 retrieval (ablation) | BM25 retriever + GPT-4o mini · NarrativeQA | ROUGE-L: 26.52 -> 30.47 | Improved |
| add HTSIR summary tree on top of SBERT retrieval (ablation) | SBERT retriever + GPT-4o mini · QASPER | Answer F1: 32.65 -> 34.48 | Improved |
| retrieve all three node layers (leaf, intermediate, root) instead of one layer (ablation) | HTSIR tree · QuALITY-HARD | accuracy: 61.54 -> 62.50 | Improved |
| retrieve two node layers instead of one (ablation) | HTSIR tree · QuALITY-HARD | accuracy: 60.57 -> 57.69 (x vs x+s) | Hurt |
| retrieve only root summary nodes (ablation) | HTSIR tree · QuALITY-HARD | accuracy: 62.50 -> 54.81 | Hurt |
| vary r (chunks per summary) from 2 to 5 (ablation) | HTSIR · QuALITY-HARD | accuracy: 62.48 -> 62.99 (r=2 to r=4) | No clear effect |
| prepend HTSIR retrieved content to full text for GPT-4o mini | GPT-4o mini · QuALITY-HARD | accuracy: 72.1 -> 74.0 | Improved |
| use HTSIR with Llama3-8B LoRA instead of plain Llama3-8B LoRA | Llama3-8B LoRA · QASPER | Answer F1: 21.7 -> 23.5 | Improved |
| use HTSIR tree instead of RAPTOR clustering tree | tree-based RAG · QASPER | Answer F1: 36.7 -> 36.9 | No clear effect |
| use HTSIR instead of graph-based RAG (HippoRAG 2) | long-document QA · NarrativeQA | F1: 25.2 -> 37.4 | Improved |
| use fixed-length chunk summaries without a tree, at tuned length (150 x 5 or 100 x 7 chunks) (ablation) | summary corpus without tree · QASPER | Answer F1 | Mixed |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- RAPTOR — RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024) · better on 3 of 3 shared results
- HippoRAG 2 — HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (NeurIPS 2024) · better on 2 of 2 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
- RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval (ICLR 2024 · similarity 0.89)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 2025 · similarity 0.89)
- QuerySum: A Multi-Document Query-Focused Summarization Dataset Augmented with Similar Query Clusters (AAAI 2024 · similarity 0.88)
- Inference Scaling for Long-Context Retrieval Augmented Generation (ICLR 2025 · similarity 0.88)
- A Theory for Token-Level Harmonization in Retrieval-Augmented Generation (ICLR 2025 · similarity 0.88)
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (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.
