Research map / Large language models / Retrieval and knowledge
When Does Divide and Conquer Work for Long Context LLM? A Noise Decomposition Framework
Zach Xu, Shang Zhu, Jue WANG, Junlin Wang, Ben Athiwaratkun, Chi Wang, James Zou, Ce Zhang
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 |
|---|---|---|---|
| split 128K KV input into 16K chunks (D&C) instead of single agent | llama70b · KV (128K) | accuracy: 0.15 -> 1.00 | Improved |
| split 128K Math input into 16K chunks (D&C) instead of single agent | llama70b · Math (128K) | accuracy: 0.09 -> 0.57 | Improved |
| split 128K KV input into 32K chunks (D&C) instead of single agent | gpt4omini · KV (128K) | accuracy: 0.60 -> 1.00 | Improved |
| choose chunk size from 3 samples instead of exhaustive search (ablation) | llama70b D&C · QA-IB (128K) | F1: 0.63 -> 0.55 | Hurt |
| choose chunk size from 5 samples instead of exhaustive search (ablation) | llama70b D&C · QA-IB (128K) | F1: 0.63 -> 0.63 | No clear effect |
| choose chunk size from 5 samples instead of exhaustive search (ablation) | qwen72b D&C · Sum (128K) | ROUGE: 0.29 -> 0.29 | No clear effect |
| add 1K token overlap between chunks (16K chunks) (ablation) | llama70b D&C · QA-IB (128K) | F1: 0.63 -> 0.54 | Hurt |
| add 1K token overlap between chunks (4K chunks) (ablation) | llama70b D&C · Sum (128K) | ROUGE: 0.23 -> 0.19 | Hurt |
| add 1K token overlap between chunks (16K chunks) (ablation) | llama70b D&C · Sum (128K) | ROUGE: 0.24 -> 0.25 | No clear effect |
| use planner-based instead of manual aggregator prompt (ablation) | D&C · Math and QA-LB (large T) | accuracy | Improved |
| increase input length for single agent from 1K to 128K | gpt4omini · KV | accuracy: 1.00 -> 0.60 | Hurt |
| increase input length for single agent from 1K to 128K | llama70b · Math | accuracy: 0.63 -> 0.09 | Hurt |
Closest papers by meaning
- Chain of Agents: Large Language Models Collaborating on Long-Context Tasks (NeurIPS 2024 · similarity 0.91)
- Training-Free Long-Context Scaling of Large Language Models (ICML 2024 · similarity 0.89)
- Structured Packing in LLM Training Improves Long Context Utilization (AAAI 2025 · similarity 0.89)
- Efficient Long Context Fine-tuning with Chunk Flow (ICML 2025 · similarity 0.89)
- Retrieval meets Long Context Large Language Models (ICLR 2024 · similarity 0.88)
- Long-Context LLMs Meet RAG: Overcoming Challenges for Long Inputs in RAG (ICLR 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.
