PaperOrchestrator: An LLM-Orchestrated multi-agent pipeline for automated end-to-end scientific paper writing

Academic paper writing is a time-consuming and demanding component of the scientific research workflow. Although large language models (LLMs) have achieved remarkable advances in text generation, single-turn direct generation strategies face critical challenges when applied to long-form, multi-section academic papers, including incomplete sections, content truncation, formatting irregularities, and reference hallucinations. We propose PaperOrchestrator, an LLM-orchestrated multi-agent pipeline system that decomposes automatic academic paper writing into seven collaborative stages: outline generation, section-by-section content generation, length-adaptive control, full-text consistency polishing, BibTeX reference generation, section-by-section robust LaTeX conversion, and journal template rendering, with inter-stage information transfer and state management achieved through a shared context data structure. On an evaluation set of 60 papers spanning natural language processing, computer vision, and biomedical AI, we compare against 11 baseline systems including GPT-4o, GPT-4-Turbo, Gemini 1.5 Pro, Claude 3.5 Sonnet (single-turn), Llama-3.1-70B, Qwen2.5-72B, and ChatPaper. PaperOrchestrator achieves state-of-the-art performance on BERTScore F1 (0.674), ROUGE-L (0.223), section completeness (96.7%), LaTeX compilation rate (90.0%), and human-evaluated overall quality (3.85/5.00, Krippendorff’s \(α =0.73\) ), with all differences reaching statistical significance after Bonferroni correction. Ablation experiments validate the necessity of core components including the Continue Mechanism and the Section-by-Section LaTeX Conversion strategy. This work provides a reproducible, systematic framework and a comprehensive empirical benchmark for LLM-driven automated academic writing.