PaperOrchestra explained: from experiment logs to a LaTeX draft
PaperOrchestra is an open-source, multi-agent paper-writing framework from researchers at Google Cloud AI Research. It takes research ideas, experimental logs and a LaTeX template, then coordinates outlining, literature review, section writing and revision, with optional plotting. It is relevant when you have research materials and need an editable first draft. The resulting claims, numbers and references still need checking before submission preparation.
The paper was submitted to arXiv on April 6, 2026. Recent discussion does not make this a November launch. The repository also states that it is not an officially supported Google product. File requirements and arguments below were checked against the official source on November 18, 2026.
What do the agents do?
| Stage | Framework task | Author check |
|---|---|---|
| Outline | Organize materials into a section plan | Does each contribution have evidence? |
| Literature review | Discover papers and check records through Semantic Scholar | Does the cited work support the claim? |
| Section writing | Draft LaTeX from the plan, materials and template | Are methods, notation and numbers faithful to the logs? |
| Refinement | Revise using simulated review feedback | Has the revision introduced errors or stronger claims? |
| Optional plotting | Generate plotting code or conceptual diagrams | Are data, axes, units and methods correct? |
Workflow illustration. Author review is an additional step recommended here. Plotting interacts with writing; this is not a complete internal execution trace.
PaperWritingBench uses materials reconstructed from 200 CVPR 2025 and ICLR 2025 papers. That supports comparisons between writing systems. Research notes from a live project can be incomplete, however, so the benchmark does not predict your draft quality or publication prospects. The project page also reports a remaining quality gap against the human-written originals.
Prepare research inputs with an evidence trail
The CLI defaults to idea_sparse.md and experimental_log.md. With plotting disabled, the code copies an existing figures/ directory and passes figures/info.json to the section writer. Prepare actual figures and description metadata compatible with your checkout; an empty folder does not establish that the pipeline will run.
raw_materials/ idea_sparse.md experimental_log.md figures/ accuracy.pdf info.jsonmy_template/ template.tex guidelines.md ... required class/style/bibliography filesThis is a directory plan, not a complete runnable sample. A custom template needs template.tex, guidelines.md and its required class, style and bibliography files. Put page limits, anonymization and figure/reference requirements in guidelines.md. The included CVPR 2025 example does not establish another venue's current requirements.
Record the question and its limits in idea_sparse.md
# Research questionWhich claim are we testing, and against which baseline?# MethodDescribe the actual algorithm, assumptions and implementation.# Evidence mapClaim C1 -> experiment E1 -> recorded metrics and figure path.# LimitsList untested settings, negative results and missing experiments.Do not fill gaps with invented results or references.Assign an evidence ID to each contribution. “Method B improves classification accuracy” needs an experiment such as E1. “The method generalizes to other datasets” needs evidence on those datasets. Keep untested ideas separate from completed contributions.
Preserve measurement conditions in experimental_log.md
# E1: illustrative data, not research findingsTask: classification on a fixed held-out splitMetric: accuracy (%)Baseline A: 82.1Method B: 83.4Difference: 1.3 percentage pointsRuns: one per method; uncertainty not estimatedInterpretation: descriptive comparison onlyNot supported: statistical significance or broad superiorityFigure: figures/accuracy.pdf (create from the recorded values)For real work also record: dataset version, split IDs, seeds,hyperparameters, code commit and raw result paths.These are illustrative numbers. An accuracy change from 82.1% to 83.4% is 1.3 percentage points. Its relative increase is approximately (83.4 - 82.1) / 82.1 × 100% = 1.58%. These descriptions are not interchangeable. One run per method also does not establish statistical significance.
Numerical audit illustration using fictional teaching data, not performance measurements of PaperOrchestra or another model.
For real experiments, record dataset versions, split IDs, seeds, hyperparameters, code commits and raw result paths. Captions should identify measurements, units and repetitions. Label conceptual diagrams separately from measured results.
What is needed to run the framework?
The official README uses Python 3.11 and requirements.txt. Configure credentials for the model services you use and for Semantic Scholar. PDF generation also depends on a working TeX toolchain and template resources; Python packages alone do not establish that all compilation dependencies are installed. Model calls can incur charges.
From the repository root, after dependencies, credentials, materials and templates are ready, the CLI invocation has this form:
python paper_writing_cli.py \ --raw_materials_dir ./raw_materials \ --latex_template_dir ./my_template \ --output_dir ./runs/draft-001 \ --idea_filename idea_sparse.md \ --experimental_log_filename experimental_log.md \ --use_plotting falseThis example disables automatic plotting to use your prepared figures. Enabling --use_plotting true also requires checking the plotting and image-model configuration. Use a fresh output directory for each draft: the current code recreates some intermediate directories, and mixing runs makes changes harder to audit.
The arguments and filenames have been checked against source. We have not executed the full model pipeline or measured its cost, speed or output quality. This is an input-preparation and workflow guide, not a verified installation walkthrough.
Check the evidence behind the draft
- Compare every table cell and numerical statement with the logs, including percentages, percentage points, means and uncertainty.
- Check training conditions, baselines and splits. Remove or flag experiments that appear only in the generated text.
- Read the cited source. Finding an API record does not establish support for the adjacent claim. Use the reference-checking workflow.
- Inspect compilation errors, unresolved citations and cross-references, and missing figures. Then review equations, captions and layout in the PDF.
- Check the target venue's anonymity, length and AI disclosure rules, and obtain author confirmation. Use pre-submission review after filling evidence gaps.
Use the preparation method without installing PaperOrchestra
If you already have a LaTeX project, organize the two research files above and ask an assistant to draft a limited section. This prompt keeps the requested scope narrow and makes missing evidence visible:
Draft only the Methods and Results sections from the supplied materials.Map every numerical claim to an experiment ID and source file.Keep the recorded metric, unit, split and number of runs unchanged.Mark missing evidence as [AUTHOR CHECK: explain what is missing].Do not invent experiments, citations or statistical significance.Return proposed LaTeX and a list of unresolved questions separately.Resolve and remove all [AUTHOR CHECK: ...] markers before submission. A prompt states requirements; you must still verify that the output follows them.
To continue editing in EasyLaTeX, organize the required .tex, .bib, style and image files in your project. Select the actual main file and the compiler required by the template. After compilation, inspect Bug / Original Log, review the PDF and download both PDF and LaTeX project for your records. See the project-to-PDF guide. This is a file handoff, not a claim that EasyLaTeX integrates PaperOrchestra or resumes its agents.