PaperOrchestra
A multi-agent framework that turns ideas and experimental logs into submission-ready LaTeX AI research manuscripts.
At a Glance
About PaperOrchestra
PaperOrchestra is a multi-agent framework from Google Cloud AI Research for automated AI research paper writing. It takes pre-writing materials such as an idea summary and experimental logs and produces LaTeX manuscripts that follow a target conference template, with literature synthesis and generated visuals. The code is released as an open-source Python project under Apache 2.0, and the README notes it is not an officially supported Google product.
What It Is
PaperOrchestra is a research-writing pipeline for researchers who have finished experiments but have not yet drafted the paper. Inputs are a directory of raw materials and a LaTeX template directory, for example a CVPR 2025 template. The output is a generated manuscript in the chosen venue format. Its authors position it as an assistive drafting tool, not an author, and say researchers remain responsible for verifying accuracy, originality and validity.
How the Multi-Agent Pipeline Works
The project page describes specialized agents that split the work:
- An Outline Agent turns the inputs into a structured plan.
- A Plotting Agent generates conceptual diagrams and statistical plots.
- A Literature Review Agent runs web searches for candidate papers and checks their existence and relevance through the Semantic Scholar API to build a citation graph.
- A Section Writing Agent writes the full LaTeX manuscript.
- A Content Refinement Agent iteratively improves the draft using simulated peer-review feedback.
Setup Path
Setup uses a Python 3.11 Conda environment and a .env file with API keys. The README lists a Semantic Scholar key and keys for Gemini (through Vertex AI or an API key) or OpenAI models. The pipeline runs through a bash script or the paper_writing_cli.py command, with options for plotting and for writer and reflection model names. The repository also includes a Streamlit demo, automated evaluation scripts and a folder for adding custom LaTeX templates.
Evaluation and Benchmark
The authors introduce PaperWritingBench, built from raw materials reverse-engineered from 200 CVPR 2025 and ICLR 2025 papers. They report that in side-by-side evaluations with 11 AI researchers, PaperOrchestra beat the baselines (a single-agent approach and AI Scientist-v2), with win margins of 50%-68% for literature review quality and 14%-38% for overall quality. They also note a quality gap remains against human-written papers. The README says the dataset is not included in the repository.
Community Discussions
Be the first to start a conversation about PaperOrchestra
Share your experience with PaperOrchestra, ask questions, or help others learn from your insights.
Pricing
Open Source (Apache-2.0)
Self-hosted PaperOrchestra core framework released under the Apache License 2.0. Users supply their own LLM and Semantic Scholar API keys (external costs).
- License: Apache License 2.0 (Apache-2.0)
- Multi-agent pipeline: Outline, Literature Review, Section Writing, Content Refinement, Plotting
- LaTeX manuscript generation using conference templates
- Python CLI and interactive Streamlit demo
- Requires own API keys (Gemini/Vertex AI/OpenAI, Semantic Scholar) - external costs
Capabilities
Key Features
- Multi-agent pipeline with outline, plotting, literature review, section writing and refinement agents
- Literature synthesis with citations verified through the Semantic Scholar API
- Plot and conceptual diagram generation
- LaTeX manuscript output following conference templates (CVPR, ICLR)
- Iterative refinement using simulated peer-review feedback
- Python CLI and bash script for running the pipeline
- Interactive Streamlit demo
- Custom LaTeX template support
- Automated evaluation scripts (autoraters)
- Support for Gemini and OpenAI models
