Google AI Research Introduces PaperOrchestra for Automated Paper Writing

Writing a research paper can be an extremely challenging process. Even after experiments are completed, a researcher faces the daunting task of translating messy lab notes, scattered results tables, and half-formed ideas into a polished, logically coherent manuscript that meets conference requirements. For many novice researchers, this translation process can feel like a graveyard for papers. A team at Google Cloud AI Research has proposed a solution called 'PaperOrchestra' — a multi-agent system that autonomously converts unstructured pre-writing materials, such as a rough idea summary and raw experimental logs, into a submission-ready LaTeX manuscript, complete with a literature review, generated figures, and API-verified citations.

Previous automated writing systems, like PaperRobot, could generate incremental text sequences but struggled to handle the full complexity of data-driven scientific narratives. More recent end-to-end autonomous research frameworks, such as AI Scientist-v1 and its successor AI Scientist-v2, automated the entire research loop but tightly coupled their writing modules to their internal experimental pipelines. You couldn't simply provide them with your data and expect a paper; they weren't standalone writers. Meanwhile, systems specialized in literature reviews, such as AutoSurvey2 and LiRA, produced comprehensive surveys but lacked the contextual awareness needed to write a targeted Related Work section that clearly positions a new method against prior art.

PaperOrchestra is specifically designed to fill this gap. It orchestrates five specialized agents that work sequentially, with two operating in parallel. The first agent generates a structured JSON outline based on the idea summary, experimental logs, and conference template. Parallel agents then execute the visualization plan and conduct a literature review, utilizing various tools to verify and confirm sources, which are subsequently used to draft different sections of the paper.

The process concludes with a section writing agent that integrates all previously generated materials and authors the remaining parts of the manuscript, such as the abstract and methodology. The final stage includes optimizing the text using a simulated peer-review system, significantly enhancing the quality of the final document. The entire pipeline requires approximately 60-70 LLM API calls and completes in an average of 39.6 minutes per paper.

Additionally, the team introduced PaperWritingBench — the first standardized benchmark specifically designed for AI research paper writing. It comprises 200 accepted papers from CVPR 2025 and ICLR 2025, allowing for testing adaptation to different conference formats.