Paper-Cleaner

Pre-submission hygiene for a research paper, in one self-contained repository:

1. Clean the LaTeX — strip a LaTeX source tree down to an arXiv-ready bundle (remove comments, dead \if…\fi branches, \todo/\note commands, resize images, compress PDFs, externalize TikZ).
2. Verify the bibliography — check that every citation is real, not retracted, not stale, and complete, against arXiv / Semantic Scholar / OpenReview / Crossref / OpenAlex.

run.sh walks you through both steps interactively.

This repo bundles two independent upstream tools as vendored source (no git submodules). Each keeps its own license — see Licensing.

Bundled tools & their original repositories

This project does not reimplement either tool; it vendors and orchestrates them. Please star / cite the upstreams — all credit for the heavy lifting is theirs:

Tool	What it does	Original repository	License
arxiv-latex-cleaner	LaTeX → arXiv-ready source	https://github.com/google-research/arxiv-latex-cleaner (Google Research)	Apache-2.0
FactReview / RefCopilot	Reference / citation fact-checker	https://github.com/DEFENSE-SEU/FactReview	AGPL-3.0

Vendored revisions: arxiv-latex-cleaner @ bcc1460 (v1.0.11-2), FactReview @ ac0f9ec (plus local modifications — see FactReview/CHANGES.md).

Quick start

# Prerequisite: uv  (https://docs.astral.sh/uv/)
#   curl -LsSf https://astral.sh/uv/install.sh | sh

git clone https://github.com/vantuan5644/Paper-Cleaner.git
cd Paper-Cleaner
./run.sh

run.sh creates one shared .venv (via uv), installs both tools editable into it, and then prompts you through cleaning and reference-checking. Press Enter to accept the [bracketed default] at each step.

LLM verifier & API keys

RefCopilot's reference verifier calls an LLM (there is no regex fallback) and benefits from a Semantic Scholar key:

# 1. An LLM provider — either Codex CLI…
npm install -g @openai/codex && codex login
#    …or set MODEL_PROVIDER + keys in FactReview/.env

# 2. Recommended: a Semantic Scholar API key (higher rate limits)
cp FactReview/.env.example FactReview/.env   # then fill in SEMANTIC_SCHOLAR_API_KEY

Without an LLM provider configured, reference checking will fail; LaTeX cleaning works regardless.

Requirements

• uv — manages the virtualenv and installs.
• Python ≥ 3.11 (required by FactReview/RefCopilot; also satisfies arxiv-latex-cleaner's ≥ 3.9).
• An LLM provider for RefCopilot (Codex CLI, or any provider configured in FactReview/.env).
• Optional: SEMANTIC_SCHOLAR_API_KEY (recommended), OPENALEX_API_KEY.
• Optional: ghostscript for PDF compression in arxiv-latex-cleaner.

Manual usage

If you'd rather drive the tools directly (after ./run.sh has built .venv, or in your own environment):

source .venv/bin/activate

# 1. Clean a LaTeX tree → produces a sibling <dir>_arXiv/
arxiv_latex_cleaner /path/to/latex --keep_bib --verbose
#    common knobs:
arxiv_latex_cleaner /path/to/latex --resize_images --im_size 500 \
                    --commands_to_delete todo note

# 2. Check references (accepts .bib | .pdf | arXiv URL | plain bib text)
refcopilot check /path/to/paper.pdf --output-dir out/
refcopilot cache prune --ttl-days 30

# 3. (Optional) apply RefCopilot's suggested BibTeX fixes back into a .bib
scripts/apply-refcheck.py out/reference_check.json refs.bib            # writes refs.bib.refcopilot
scripts/apply-refcheck.py out/reference_check.json refs.bib --in-place # overwrite (backup at .bak)

See arxiv_latex_cleaner --help and refcopilot check --help for the full flag sets. While iterating, RefCopilot's --no-llm-verify, --no-cache, and --max-refs N are the handy ones.

Repository layout

Paper-Cleaner/
├── run.sh                  # interactive driver (MIT)
├── scripts/
│   ├── apply-refcheck.py   # apply RefCopilot BibTeX fixes to a .bib (MIT)
│   └── revendor.sh         # re-vendor a tool from a pinned upstream commit (MIT)
├── arxiv-latex-cleaner/    # vendored — Apache-2.0 (unmodified)
├── FactReview/             # vendored — AGPL-3.0 (modified; see CHANGES.md)
│   └── RefCopilot/         #   the reference checker used here
├── LICENSE                 # MIT — scope note + the top-level glue's terms
└── NOTICE                  # per-component attribution + AGPL §13 notice

The glue is deliberately loose coupling: run.sh only invokes the two CLIs as subprocesses and apply-refcheck.py only reads RefCopilot's JSON output — the two tools never import each other.

Licensing

This repository is an aggregate of independently licensed components. There is no single license for the whole repo.

Path	License	Notes
LICENSE, run.sh, scripts/, docs	MIT	Top-level glue authored here
arxiv-latex-cleaner/	Apache-2.0	© 2018 The Google Research Authors; unmodified
FactReview/ (incl. RefCopilot/)	AGPL-3.0	© FactReview authors; modified (Crossref backend, FactReview/CHANGES.md)

Why these can be bundled: Apache-2.0 is one-way compatible into AGPL-3.0, and the two tools are run as separate programs (no cross-imports), so this is a "mere aggregation" — each component keeps its upstream license and the MIT glue does not relicense anything.

AGPL-3.0 network-use notice (§13): RefCopilot is AGPL-3.0 and is modified here. If you run a modified version to interact with users over a network, you must offer those users the Corresponding Source of your version. The complete source bundled here is this repository.

Full texts: arxiv-latex-cleaner/LICENSE, FactReview/LICENSE, and NOTICE.

Updating a vendored tool

Vendored trees are pinned by commit (recorded in NOTICE). To pull a newer upstream revision into a subtree:

scripts/revendor.sh arxiv-latex-cleaner   # re-vendor from its recorded upstream
scripts/revendor.sh FactReview <commit>   # pin a specific commit

After re-vendoring, update the pinned commit in NOTICE, and if you re-vendor FactReview, re-apply / re-document any local modifications in FactReview/CHANGES.md.

Contributing

• Changes to a vendored subtree must respect that subtree's upstream license. Modifications to FactReview/ (AGPL-3.0) must be recorded in FactReview/CHANGES.md with a date (AGPL §5a).
• Never add an import that crosses between arxiv-latex-cleaner/ and FactReview/ — that would break the aggregation boundary and pull the Apache-2.0 code under AGPL-3.0.
• Run each tool's own test suite before submitting (see below).

Tests

source .venv/bin/activate
python -m unittest arxiv_latex_cleaner.tests.arxiv_latex_cleaner_test   # arxiv-latex-cleaner
( cd FactReview/RefCopilot && pytest )                                  # RefCopilot