google-research/paper-orchestra
AI EngineeringTurns research ideas and experiment logs into LaTeX papers through a multi-agent writing pipeline.
What it does
PaperOrchestra is a Python framework from Google Cloud AI Research that turns unstructured pre-writing materials—an idea and an experimental log—into a submission-ready LaTeX manuscript. A pipeline of specialized agents handles outline generation, literature search and synthesis, section writing, content refinement, and plotting.
The literature agent searches Semantic Scholar and integrates results into the draft, while the plotting agent can generate figures from code. The output follows a supplied conference LaTeX template. The project targets researchers who have results and notes but want help drafting a paper; it is not a model release or a hosted writing service.
Getting started
You need Conda, Python 3.11, and API credentials for the model providers you plan to use.
conda create -n paper_orchestra python=3.11
conda activate paper_orchestra
pip install -r requirements.txt
Create a .env file in the repository root with the keys for your setup: SEMANTIC_SCHOLAR_API_KEY for literature search, and either GEMINI_API_KEY or VERTEX_AI_PROJECT plus VERTEX_AI_LOCATION for Gemini, or OPENAI_API_KEY for OpenAI models. Optional SMTP_EMAIL and SMTP_PASSWORD fields enable email delivery at the end of the demo run.
Then run the pipeline with your materials, a LaTeX template directory, and an output directory:
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
--use_plotting defaults to false; enabling it activates the plotting agent. --writer_model_name and --reflection_model_name select the LLMs used by the writing/literature and reflection agents. A shell wrapper, ./paper_writing_cli.sh, is also provided. An interactive Streamlit demo lives in frontend/, with run instructions in its own frontend guide.
Constraints
The research dataset used for generation is not included and was promised for a later release; until then, you must supply your own raw materials. Agents require configured API keys and may incur provider usage. Output quality depends on the supplied template and materials, and template support is extended by adding subfolders under templates/. The repository is Apache-2.0 licensed, and the README states it is not an officially supported Google product. The plotting workflow only applies when explicitly enabled.
HotRepo's take
On 2026-11-16 the repository recorded 1,024 stars and 169 forks, with 306 stars added in that daily window and a best all-language daily rank of 12. That is a single snapshot, so it supports one day of attention rather than a trend. The premise—automating literature synthesis and LaTeX assembly while keeping the researcher's raw materials as input—is well matched to the multi-agent design, and the missing dataset release is the main gap to watch for readers who want to reproduce the full pipeline rather than draft their own notes.
For project background and links, see the project page; the code is in the PaperOrchestra repository.
Written by AI from the project's README, updated . The project's own docs are the source of truth.