Google Research Releases PaperOrchestra: Five AI Agents Write a Research Paper from Drafts
You have the idea and the experiment logs, but no paper: you still need an outline, a review of prior work, figures, and typesetting. Google Research has released PaperOrchestra, five AI agents that assemble a LaTeX manuscript in the format used by scientific conferences from raw drafts. One agent builds the outline, another searches the literature, and the rest write sections, revise the text, and draw the figures. The repository gained 306 stars in a day, but no quality figures have been published yet.
AI-processed from GitHub Trending AI; edited by Hamidun News
Picture a PhD student whose experiments have already run: there are text notes with the idea and tables with the results of each run. Even so, weeks of routine work remain before a finished paper: building the structure, finding and summarizing other people's work, drawing the charts, and then fitting everything to the conference requirements. Google Research has published PaperOrchestra on GitHub, a project that tries to automate precisely this part.
It runs five AI agents (separate instances of a language model, each with a narrow role), and the output is a manuscript in LaTeX, the typesetting language in which scientific conferences accept papers. In a single day the repository gained 306 stars, meaning 306 GitHub users marked it as interesting, and it landed on the trending list.
What PaperOrchestra does
It is a multi-agent framework: instead of one large model that is told "write a paper," a team of specialists works together. According to the authors, these are a planning agent, a literature review agent, a section-writing agent, a text-refinement agent, and a figure-building agent.
The input is raw material: a description of the idea and an experiment log. No strict structure is required. The output is a manuscript formatted to the target conference's template, with a literature review and generated illustrations: result plots and conceptual diagrams.
If you break down a PhD student's work by role, you get exactly this scheme. The first agent sketches the paper's skeleton. The second searches for relevant work, selects the most important papers, and weaves them into the text. The third writes the sections. The fourth rereads the text and improves it. The fifth builds the plots from code.
How to run it
The project installs into a separate Conda environment on Python 3.11, with dependencies taken from requirements.txt. You need keys for whichever models you plan to use. For Gemini, you need either a project and a Vertex AI region or GEMINI_API_KEY; for OpenAI models, OPENAI_API_KEY. A Semantic Scholar key is specified separately, as it is needed for literature search. If you want the demo to send the result to your email, two more fields are added.
You launch it with the paper_writing_cli.py script. There are two required parameters: a folder with the source materials and a folder with the LaTeX template. The output folder can be omitted, in which case a timestamped directory is created. There is a switch for the figure agent and a separate choice of model for the writing and literature agents, and for the reviewer agents that check and rethink the draft. For those who would rather not work in a console, there is an interactive Streamlit demo.
The repository contains automatic evaluation scripts, the core logic of the agents and the pipeline, helper functions for API calls and PDF handling, and conference templates. You can add your own template by creating a subfolder.
Limitations and what to do about them
It's important to look soberly at what the material does not contain. There are no quality figures: it is unclear how often the generated papers pass peer review or how accurate the literature citations are. The dataset on which the project was tested has not been published yet; the authors promise to release it later. The work is described in the preprint arXiv:2604.05018 (Yiwen Song, Yale Song, Tomas Pfister, Jinsung Yoon). The repository is explicitly marked as an unofficial Google product, so do not expect support or guarantees.
There is also an ethical question that the code does not resolve. Automatically assembling a manuscript from logs can easily turn into a flood of weak papers, and many conferences have rules on the use of AI in writing. Responsibility for the accuracy of facts, the correctness of references, and the honest description of experiments remains with the author.
The practical takeaway for agent developers: PaperOrchestra is a clear example of an architecture where a task is split into narrow roles and a separate self-review stage follows the writing. This pattern can be carried over to other long documents: technical reports, documentation, analytical memos. For researchers, it is a way to quickly get a first draft and a base for edits, but not a replacement for their own work. Until the dataset and independent evaluations are out, the project should be treated as an interesting starting point, not a ready tool for submitting papers.
- Nov 21, 2026Grok Build left 35,578 apps open — 94.7% of live projects require no login
- Nov 21, 2026Signal Blocks Screenshots on iOS 27 and Android to Keep AI Agents From Reading Chats
- Nov 21, 2026Hassabis's Isomorphic Labs Raises $5 Billion for AI Drug Design
- Nov 21, 2026Google gives Claude and Cursor a DevTools panel: the agent finds out for itself why a site is slow
Key Questions Answered
What PaperOrchestra does?
It is a multi-agent framework: instead of one large model that is told "write a paper," a team of specialists works together. According to the authors, these are a planning agent, a literature review agent, a section-writing agent, a text-refinement agent, and a figure-building agent. The input is raw material: a description of the idea and an experiment log. No strict structure is required.
How to run it?
The project installs into a separate Conda environment on Python 3.11, with dependencies taken from requirements.txt. You need keys for whichever models you plan to use. For Gemini, you need either a project and a Vertex AI region or GEMINI_API_KEY; for OpenAI models, OPENAI_API_KEY. A Semantic Scholar key is specified separately, as it is needed for literature search. If you want the demo to send the result to your email, two more fields are added. You launch it with the paper_writing_cli.py script.
What are the key details on Limitations and what to do about them?
It's important to look soberly at what the material does not contain. There are no quality figures: it is unclear how often the generated papers pass peer review or how accurate the literature citations are. The dataset on which the project was tested has not been published yet; the authors promise to release it later. The work is described in the preprint arXiv:2604.05018 (Yiwen Song, Yale Song, Tomas Pfister, Jinsung Yoon).
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.