opensourceprojects.dev

A broadsheet for software that doesn't ask for your email

Multi-agent pipeline that turns research notes into LaTeX manuscripts
GitHub RepoViews8

Project Description

View on GitHub

PaperOrchestra: A Multi-Agent Pipeline That Turns Your Research Notes Into a LaTeX Paper

You've got the experiments logged, an idea that seems promising, and a pile of notes sitting in a directory somewhere. What you don't have is a draft. Writing the paper is the part that eats your nights—outlining, chasing down related work, writing each section, then wiring it all into a conference template. PaperOrchestra is an open-source framework from Google Cloud AI Research that takes those rough pre-writing materials and runs them through a coordinated set of agents to produce a submission-ready LaTeX manuscript.

What It Does

PaperOrchestra is a multi-agent framework for automated AI research paper writing. You feed it unconstrained pre-writing materials—ideas and experimental logs—and it transforms them into a full LaTeX manuscript, complete with literature synthesis and generated visuals like plots and conceptual diagrams.

The architecture is the interesting part. Rather than one monolithic prompt, it acts like an orchestrated team of specialized agents. There's an Outline agent, a Literature Review agent, a Section Writing agent, a Content Refinement agent, and a Plotting agent. Each handles its own piece of the job, moving from raw materials to a formatted paper. The output adheres to your target conference template, so you're not hand-fixing formatting after the fact. It's built in Python (set up with a Conda environment on Python 3.11), driven by a CLI, and it works with multiple model providers—Gemini via Vertex AI or an API key, and OpenAI models.

Why It's Cool

  • The division of labor is the real idea here. A lot of "write my paper" tools are just a single model asked to do everything at once. PaperOrchestra splits the work into specialized stages—outline, literature, writing, refinement, plotting—which mirrors how people actually write papers. Each agent gets a narrower job, which is a more sensible design than one giant generation pass.

  • Literature synthesis is built in, not bolted on. It automatically searches, prioritizes, and integrates relevant literature into the manuscript. Anyone who's spent a weekend chasing citations will see the appeal. It uses the Semantic Scholar API for this, which is a key you'll need to supply.

  • It generates figures from code. With the plotting workflow enabled, it produces plots and supports including conceptual diagrams. That's a step beyond text generation—figures are often the most tedious part to get right.

  • Template-driven output. You point it at a LaTeX template directory (the README references templates/cvpr2025 as an example), and it produces a manuscript that conforms to that template. Submission-ready formatting without manual cleanup is a practical win.

  • Model flexibility. You can separately choose the model for writer and literature agents versus reflection agents, which means you can mix providers based on cost or capability.

How to Try It

Start by cloning the repo and setting up the environment:

conda create -n paper_orchestra python=3.11
conda activate paper_orchestra
pip install -r requirements.txt

Then create a .env file in the project root with your API keys. You'll need a SEMANTIC_SCHOLAR_API_KEY for the literature side, plus keys for whichever model provider you're using—either VERTEX_AI_PROJECT and VERTEX_AI_LOCATION (or GEMINI_API_KEY) for Gemini, or OPENAI_API_KEY for OpenAI. There are optional SMTP fields if you want the demo to email you the result.

Run the pipeline via the bash script:

./paper_writing_cli.sh

Or call the Python CLI directly:

python paper_writing_cli.py \
  --raw_materials_dir /path_to_your_raw_materials/ \
  --latex_template_dir /path_to_your_latex_template/ \
  --output_dir /path_to_your_output_folder/ \
  --use_plotting true

The required arguments are --raw_materials_dir and --latex_template_dir. If you skip --output_dir, it creates a timestamped folder for you. Note that --use_plotting defaults to False, so you'll want to flip it on if you want figures. One heads-up: the dataset used for paper generation isn't included in the repo yet—the README says it'll be released separately, so you'll be working with your own materials for now.

You can find everything at github.com/google-research/paper-orchestra.

Final Thoughts

PaperOrchestra is best suited to researchers and grad students who already have the substance of a paper—results, logs, a clear idea—and want help turning it into a structured draft. It's not going to invent your contributions, and the missing dataset means you're bringing your own inputs. But the multi-agent split and the built-in literature and plotting stages make it a genuinely interesting take on the drafting problem. If you've been sitting on a folder of results waiting for the writing to happen, this is worth a look.


Follow @githubprojects for more developer tools and open source projects.

Back to Projects
Project ID: paper-orchestraLast updated: November 18, 2026 at 06:12 AM