Google Research has released PaperOrchestra, a multi-agent framework that automates the transformation of raw research materials — ideas and experimental logs — into submission-ready LaTeX manuscripts. The system orchestrates five specialized agents: Outline, Literature Review, Section Writing, Content Refinement, and Plotting, which collectively handle end-to-end paper generation. The Literature Review agent automatically searches, prioritizes, and integrates relevant citations, while the Plotting agent generates data visualizations and supports conceptual diagrams. Output adheres to target conference templates, with CVPR 2025 provided as an example. The pipeline requires Python 3.11 in a Conda environment and depends on external API keys for LLM providers including OpenAI (OPENAI_API_KEY), Google Gemini (GEMINI_API_KEY or Vertex AI project/location), and Semantic Scholar (SEMANTIC_SCHOLAR_API_KEY). Execution is driven via a CLI or bash script with required arguments for raw materials and LaTeX template directories, optional output directory, and a flag to enable plotting (--use_plotting true). Separate model parameters control the writer/literature agents (--writer_model_name) and reflection agents (--reflection_model_name). An interactive Streamlit demo is included in the frontend directory. The repository structure separates core agent logic (methods/), evaluation scripts (autoraters/), LaTeX templates (templates/), and utilities (utils/). The training dataset is not yet released. The work is cited as an arXiv preprint (arXiv:2604.05018) and is not an officially supported Google product.
Read original
github-trending/python