Research map / Multi-agent systems and games / LLM agents
DeepWriter: A Multi-Agent Collaboration Framework for Information-rich Ultra-long Book Writing
Ming Wang, Minghao Hu, Xiuli Kang, Li He, Yu Tian, Chunming Liu, Han Shi, Zhunchen Luo, Wei Luo, Guotong Geng
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeepWriter, a multi-agent collaboration framework for generating information-rich, ultra-long books (100k+ words).
Problem
Existing long-text generation methods cannot produce ultra-long, book-scale content and lack mechanisms to integrate rich factual information.
Approach
A multi-agent collaboration framework for book writing that targets both book-scale length and integration of factual information into the generated text.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add OutlineBot planning and retrieval-augmented WriteBot (ablation) | Direct Gen · DeepWriter-Bench | BookScore: 31.93 -> 58.35 | Improved |
| add TriviaBot (ablation) | RAG Gen · DeepWriter-Bench | BookScore: 58.35 -> 67.54 | Improved |
| add ImageBot (ablation) | DeepWriter Trivia · DeepWriter-Bench | BookScore: 67.54 -> 80.92 | Improved |
| replace Co-STORM with full DeepWriter | Co-STORM · DeepWriter-Bench | BookScore: 47.23 -> 80.92 | Improved |
| use OutlineBot expert-panel deliberation instead of Co-STORM | Outline generation · DeepWriter-Bench | heading soft recall: 79.65 -> 89.37 | Improved |
| use OutlineBot instead of Direct Gen outline | Outline generation · DeepWriter-Bench | heading entity recall: 31.26 -> 62.12 | Improved |
| use WriteBot instead of Co-STORM | Main text · DeepWriter-Bench | ROUGE-1: 45.32 -> 60.34 | Improved |
| use WriteBot instead of Co-STORM | Citation verifiability · DeepWriter-Bench | citation recall: 65.73 -> 77.35 | Improved |
| add context-aware retrieval for trivia (TriviaBot vs Trivia RAG) (ablation) | Trivia generation · DeepWriter-Bench | factual accuracy: 4.03 -> 4.63 | Improved |
| add retrieval to trivia generation (Trivia RAG vs Trivia Direct) (ablation) | Trivia generation · DeepWriter-Bench | factual accuracy: 3.41 -> 4.03 | Improved |
| use ImageBot instead of Image FuzzyCrawl (ablation) | Image insertion · DeepWriter-Bench | semantic consistency: 3.45 -> 4.32 | Improved |
Closest papers by meaning
- Agents' Room: Narrative Generation through Multi-step Collaboration (ICLR 2025 · similarity 0.88)
- LongWriter: Unleashing 10,000+ Word Generation from Long Context LLMs (ICLR 2025 · similarity 0.85)
- WritingBench: A Comprehensive Benchmark for Generative Writing (NeurIPS 2025 · similarity 0.84)
- BooookScore: A systematic exploration of book-length summarization in the era of LLMs (ICLR 2024 · similarity 0.84)
- LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs (ICLR 2025 · similarity 0.84)
- Chain of Agents: Large Language Models Collaborating on Long-Context Tasks (NeurIPS 2024 · similarity 0.84)
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.
