arXiv source bundle
The LaTeX source bundle arXiv expects you to upload — pandoc has already been run, the three files inside are what arXiv's TeX Live builds the PDF from.
How to submit
- On arXiv’s submission form, choose the “upload a single
.tar.gz/.zip” option. - Upload the file linked above.
- On the Select Compiler screen, keep the auto-selection:
compiler
pdflatex, TeX Live 2025, top-levelpaper.tex. All three files (paper.tex,paper.tex.body,neurips_2026.sty) should be marked Keep. - arXiv runs
pdflatex(notpandoc— pandoc was already run when the bundle was built) and produces the public PDF. - In the Comments metadata field, paste the AI-disclosure one-liner: AI language models were used for literature exploration and drafting assistance; experiments, results, and final text are the author’s own. Full disclosure in the AI-use statement.
What’s inside, and how it builds
The build chain is:
paper.md ─pandoc→ paper.tex.body ─\input→ paper.tex ─pdflatex→ paper.pdf
↑
neurips_2026.sty
arXiv’s side only runs the second half (pdflatex paper.tex against
the three files in the tarball), so the bundle ships pandoc’s output
pre-generated. Three flat files, no nesting:
paper.tex— the wrapper: documentclass, packages, hyperref config, Unicode → glyph mapping, the pandoc-longtable + hyperrefnone-counter shim, title and author block,\input{paper.tex.body}.paper.tex.body— the entire paper body as LaTeX, produced by pandoc frompaper/paper.mdon every CI run.neurips_2026.sty— the official NeurIPS 2026 style file, loaded inpreprintmode (named author block, no line numbers).
Reproducibility
The replication package referenced in the paper’s Reproducibility
statement is at /sutra-replication-package.zip.
You can attach it as an arXiv ancillary file alongside this source
bundle, or rely on the in-paper GitHub URL. The arXiv-hosted ancillary
is the durable copy if the GitHub repository ever moves.
This page and the tarball are regenerated by the
Pages deploy workflow
on every change to paper/.