Rank #01 · Daily · AI
PaperOrchestra: Google Cloud AI Research's Multi-Agent Paper-Writing Framework
Point the pipeline at your idea notes, experimental logs, and a conference LaTeX template, and specialized agents draft a manuscript — literature, sections, refinement, and optional plots included.
What it does
PaperOrchestra is a multi-agent framework for automated AI research paper writing, published under google-research by authors at Google Cloud AI Research (Yiwen Song, Yale Song, Tomas Pfister, and Jinsung Yoon). GitHub description: "PaperOrchestra is a multi-agent framework for automated AI research paper writing." Project page: yiwen-song.github.io/paper_orchestra. Companion preprint: arXiv:2604.05018. The README disclaimer states this is not an officially supported Google product.
README, LICENSE, and the official project page (as verified):
- The pipeline is designed to turn unconstrained pre-writing materials — idea summaries and experimental logs — into submission-ready LaTeX manuscripts, including literature synthesis and generated visuals (plots and conceptual diagrams).
- Specialized agents cover outline generation, literature review, section writing, content refinement, and plotting. The project page describes the Outline Agent synthesizing inputs into a plan; the Literature Review Agent searching the web and checking candidates via the Semantic Scholar API; the Section Writing Agent drafting the LaTeX; the Content Refinement Agent iterating with simulated peer-review feedback; and the Plotting Agent producing diagrams and statistical plots.
- LaTeX automation targets conference templates you supply (the repo ships examples under
templates/, such astemplates/cvpr2025; you can add your own subfolder). The project page shows gallery PDFs in CVPR and ICLR layouts, both with ground-truth visuals and with synthesized visuals. - Install and run: a Conda environment on Python 3.11, then
pip install -r requirements.txt. Configure API keys in a.envfile (Semantic Scholar; OpenAI and/or Vertex AI / Gemini, depending on the models you choose). Optional SMTP settings are only for the demo's email step. Run./paper_writing_cli.shorpython paper_writing_cli.pywith--raw_materials_dir,--latex_template_dir, optional--output_dir, and--use_plotting(defaultFalse). Writer and reflection model names are also configurable. A Streamlit interactive demo lives underfrontend/. - The evaluation dataset is not included in this repository and is described as being released later. The project page calls that benchmark PaperWritingBench: reverse-engineered raw materials from 200 papers (100 each from CVPR 2025 and ICLR 2025). Treat the dataset description and any later release date as the authors' plan, not a TidyHow delivery promise.
- The project page reports author-run human evaluations (11 AI researchers, side-by-side, blinded) claiming win-rate margins over a single-agent baseline and AI Scientist-v2, and a remaining quality gap versus human-written ground truth. Those figures are the paper's results, not a TidyHow evaluation. Do not read them as conference acceptance rates.
- An ethics statement on the project page positions the system as an assistive drafting tool, not an independent author: researchers remain accountable for accuracy, originality, and validity. API-grounded citation checks are described as a safeguard, not a substitute for human review.
License: Apache-2.0 (LICENSE file and GitHub SPDX agree).
Who it's for
- AI researchers who already have idea notes and experimental logs and want help turning them into a conference-style LaTeX draft.
- Labs exploring a multi-agent writing pipeline with literature search, optional plot generation, and template-based compilation.
- Readers of the PaperOrchestra paper who want to inspect or rerun the open-source code.
Less ideal if you expect a system that can replace a researcher, submit a paper without human review, or invent results you did not record in the logs. It is also a poor fit if you cannot supply model and Semantic Scholar API credentials, need the evaluation dataset in-tree today (it is not in this repo), or want a general-purpose writing assistant outside this LaTeX / AI-conference workflow.
Stars / language / license
4,288 stars · Python · Apache-2.0
Data verified: 2026-11-19 09:04 Asia/Shanghai (official Daily Top10 gate; language/license confirmed via GitHub API get_repository + README + LICENSE + official project page)
Trending note: +1,524 stars today (captured 09:04 Asia/Shanghai)
Category tags
AI · multi-agent · paper-writing · LaTeX · Python · Apache-2.0 · Google-Cloud-AI-Research · literature-review
How it differs from similar projects
Compared with asking a general LLM to "write the paper." PaperOrchestra's pitch is a staged, specialized-agent pipeline around your raw materials and a real conference template, plus literature lookup grounded in an external paper API. That is an architecture claim, not a quality guarantee.
Compared with end-to-end "AI scientist" systems that also run experiments. This repository is framed as a writing framework: it starts from pre-writing materials you already have. It does not claim to design, run, or validate the underlying experiments.
Compared with writing the paper yourself in Overleaf or a local TeX tree. Those remain the usual path to a submission. PaperOrchestra is a drafting accelerator; the authors say a quality gap versus human-written ground truth remains, and humans stay accountable for the manuscript.
Caveats. Human-evaluation margins and baseline comparisons are author-reported. No conference acceptance rate, novelty score, or independent audit is claimed here. Generated citations, numbers, and plots still need a researcher's review. The evaluation dataset is not in the repo. This is not an officially supported Google product.