Chapter 9: Reading Papers Like Source Code, and a Cheat Sheet

The fastest way to get from intermediate to fluent: read the LaTeX source of papers you admire, next to their PDFs. arXiv makes this possible for most papers.

9.1 Getting the source of any arXiv paper

1. Open the abstract page, e.g. https://arxiv.org/abs/1706.03762.
2. Under Access Paper, click “TeX Source” (or Other formats → Download source). The direct URL pattern is https://arxiv.org/src/1706.03762 (older pattern: /e-print/).
3. You get a gzipped tarball (sometimes a single gzipped .tex):

mkdir attention && cd attention
curl -L -o src.tar.gz https://arxiv.org/src/1706.03762
tar -xzf src.tar.gz || (gunzip -c src.tar.gz > main.tex)
ls

4. Find the main file: grep -l '\\documentclass' *.tex.
5. Compile it locally (latexmk -pdf main.tex) or upload the folder to Overleaf as a zip. Then change things and watch what happens.

Authors’ comments are often still in there. It’s a fascinating look at how papers get written, and sometimes at what got cut.

9.2 Working backwards: “how did they make that?”

9.3 A reading exercise

Pick three papers from arXiv’s cs.LG listing:

1. A theory paper: look at their \newtheorem setup, macros, how proofs are split into lemmas, restatable usage.
2. An empirical / benchmark paper: look at the table code, how they generate tables, \resizebox usage, colors.
3. A systems / architecture paper: look at the TikZ or PDF figures, algorithm boxes, wrapfigure layout.

For each paper, write down:

• Preamble packages you haven’t seen → look them up with texdoc <pkg>.
• Their macro file: steal good ideas for your own macros.tex.
• One trick you didn’t know.

Then rewrite one page of each paper from the PDF alone, without looking at the source, and compare.

9.4 Common pitfalls (and fixes)

9.5 Typst equivalents (for when you try it)

9.6 One-page cheat sheet

%================= PREAMBLE =================
\documentclass[11pt]{article}            % or the venue template
\usepackage[T1]{fontenc}
\usepackage{microtype}
\usepackage{amsmath,amssymb,amsthm,mathtools,bm}
\usepackage{graphicx}                    \graphicspath{{figures/}}
\usepackage{subcaption,booktabs,multirow,siunitx}
\usepackage{xcolor}
\usepackage{enumitem} \setlist{nosep}    % compact lists everywhere
\usepackage{algorithm,algpseudocode}
\usepackage[numbers]{natbib}             % templates usually set this
\usepackage{hyperref}                    % late
\usepackage[capitalize,noabbrev]{cleveref} % after hyperref
\input{macros}

%================= STRUCTURE =================
\title{} \author{} \maketitle
\begin{abstract} \end{abstract}
\section{} \subsection{} \paragraph{Bold.} \section*{No number}
\appendix
\input{sections/intro}

%================= TEXT =================
\textbf{} \emph{} \texttt{} \textsc{} \url{} \href{url}{text}
\footnote{}   ~ (nbsp)   -- (range)   --- (dash)   ``quotes''
\% \$ \& \# \_ \{ \}
\begin{itemize} \item \end{itemize}     \begin{enumerate} \item \end{enumerate}

%================= MATH =================
$inline$   \[ display \]   \begin{equation} \label{eq:x} \end{equation}
\begin{align} a &= b \\ &= c \nonumber \end{align}
\frac{}{} \sqrt{} \sum_{i}^{n} \int_a^b \prod \lim_{x\to0}
\mathbb{R} \mathcal{L} \bm{x} \mathbf{W} \mathrm{d}x \text{if}
\left( \right)  \bigl( \bigr)  \mid  \top  \odot  \coloneqq
\begin{bmatrix} a & b \\ c & d \end{bmatrix}   \begin{cases} \end{cases}
\DeclareMathOperator*{\argmin}{arg\,min}
\begin{theorem} \end{theorem}  \begin{proof} \end{proof}

%================= FLOATS =================
\begin{figure}[t] \centering \includegraphics[width=\linewidth]{f}
  \caption{} \label{fig:f} \end{figure}
\begin{table}[t] \centering \caption{} \label{tab:t}
  \begin{tabular}{@{}lcc@{}} \toprule A & B & C \\ \midrule ... \\ \bottomrule \end{tabular}
\end{table}

%================= REFERENCES =================
\cref{fig:f} \Cref{eq:x} \eqref{eq:x} \citet{k} \citep{k}
\bibliographystyle{plainnat} \bibliography{refs}

%================= BUILD =================
% latexmk -pdf main.tex   |   latexmk -pdf -pvc main.tex   |   latexmk -C

9.7 Where to go next

• Overleaf’s “Learn LaTeX in 30 minutes” and its documentation pages, which are excellent per-topic references.
• “The Not So Short Introduction to LaTeX2ε” (texdoc lshort): the classic free book.
• “LaTeX for Beginners” / learnlatex.org: interactive lessons that run in the browser.
• Package manuals via texdoc: amsmath (texdoc amsldoc), booktabs, siunitx, cleveref, tikz, biblatex.
• TeX StackExchange for every “how do I…” question.
• The LaTeX Companion, 3rd ed. (Mittelbach & Fischer, 2023): the definitive reference book.
• Read arXiv sources of 10 papers in your field. This beats every tutorial.

9.8 Final project

Take example-paper/ and turn it into a 4-page mock paper on a topic you know:

1. Swap in the current NeurIPS or ICML style file (download it from the CFP) and adapt the author block.
2. Write your own macros.tex notation.
3. Add a matplotlib-generated PDF figure at the correct width.
4. Generate one results table from a Python script.
5. Add a theorem with a proof in the appendix using restatable.
6. Run arxiv_latex_cleaner and check that the cleaned version compiles with just latexmk -pdf.

If you can do all six without searching more than a couple of times, you’re past intermediate.