Google recently released PaperOrchestra (arXiv:2604.05018), a multi-agent framework that converts unstructured research materials, such as logs, ideas, and results, into submission-ready LaTeX manuscripts.
It employs a specialized 5-aagents pipeline: Outline, Plotting/Lit Review, Section Writing, and Refinement.
This setup greatly surpasses single-agent models in literature review quality and overall performance.
I created this repository to transform the paper’s prompts, schemas, and verification gates into a "skill pack" that any modern coding agent can use.
Repo: https://github.com/Ar9av/paper-orchestra
I am thinking of improving on it through:
- optional semantic scholar support for verifying
- an arxiv packager that strips comments and zips everything up for submission in one click.
- human-in-the-loop checkpoints that pause the pipeline so you can approve the outline before it starts burning tokens
