Wink Pings

Google open-sources an AI band that writes papers; reviewers check citations first

Google Cloud AI Research has open-sourced PaperOrchestra, which breaks paper writing into several agents for outlining, literature review, section writing, content polishing, and figure generation. You input ideas and experiment records, and it outputs a submittable LaTeX manuscript. Some praise it, while others warn: audit the citations first, then talk about submitting.

Google Cloud AI Research has split paper writing into a multi-agent band, and the code is already open source. The project is called PaperOrchestra: you input ideas and experiment records, and it outputs a submittable LaTeX manuscript. The repository states it clearly: it uses a set of specialized agents to handle outline generation, literature review, section writing, content polishing, and figure generation, then fits everything into the target conference's LaTeX template.

PaperOrchestra 架构总览

Several stages in the pipeline are worth examining separately. The literature agent searches, ranks, and synthesizes relevant papers; the writing agent expands each section according to the outline; the content-polishing agent handles review and rewriting; the plotting agent generates charts from code and can also insert conceptual diagrams. For people racing toward a deadline, the most immediate convenience is LaTeX automation—no need to wrestle with formatting and templates from scratch.

Getting started is not exactly plug-and-play. The repository requires Python 3.11 and a Conda environment, and you need to configure API keys for Semantic Scholar, OpenAI, or Gemini/Vertex. The entry points are paper_writing_cli.sh or python paper_writing_cli.py, and the core parameters include raw_materials_dir, latex_template_dir, output_dir, and use_plotting. The repository also includes an interactive Streamlit demo. The dataset was not released with the code; the team says it will be released separately later.

Some people in the replies were blunt. Julian Scott said that figure generation and section drafting do save time, but the citation agent is the first thing he would audit; fabricated references appearing at scale are more troublesome than poor writing. Techboy added that writing has never been the hardest part—what is useful is confirming that citations actually exist. A cleanly formatted LaTeX manuscript that cites nonexistent papers is worse than a messy draft. Hakimi wondered whether citations would hold up on obscure topics, while KRSNA felt that an open-source citation agent at least lets people inspect what it feeds into the manuscript; the review-rewrite loop is the part to watch. Someone else asked: once the agent finishes writing, who counts as first author?

PaperOrchestra compresses the paper-production process into a pipeline and shifts the bottleneck to verification. Open-sourcing the code lets people inspect the citation chain, but whether the citations are real, whether obscure fields are covered, and whether generated figures match the experiments still require human oversight. Don't hand over the submit button just yet. It is more like a research assistant; the corresponding author is still human.

Repository: https://github.com/google-research/paper-orchestra

Original post: https://x.com/RoundtableSpace/status/2122627051517833466

Google's disclaimer states that the project is not officially supported.

发布时间: 2026-11-17 23:37