ML Conference Paper (NeurIPS / ICML / ICLR style) LaTeX Template

% ─────────────────────────────────────────────────────────────────────────
% ML CONFERENCE PAPER TEMPLATE
% Works for NeurIPS / ICML / ICLR / CVPR / AAAI submissions.
%
% To use the official style file:
%   NeurIPS  → \usepackage{neurips_2025}
%   ICML     → \usepackage{icml2025}
%   ICLR     → \usepackage{iclr2025_conference}
% Download from the conference website and place next to this .tex file.
% ─────────────────────────────────────────────────────────────────────────
\documentclass{article}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage[margin=2.5cm, top=2.5cm]{geometry}
\usepackage{amsmath,amssymb,amsthm}
\usepackage{graphicx}
\usepackage{booktabs}
\usepackage[hidelinks]{hyperref}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{microtype}
\usepackage{xcolor}
\usepackage{natbib}

% ── Anonymous submission header (remove for camera-ready) ───────────────
\title{\textbf{Title of Your Paper}\\
  \large Submitted to [Conference Year] --- Under Review}

% For camera-ready, replace with real author block:
% \author{
%   First Author\thanks{Equal contribution.} \And Second Author \AND Third Author \\
%   University Name \quad\quad Lab / Affiliation \\
%   \texttt{\{first,second\}@university.edu}
% }
\author{Anonymous Authors}
\date{}

% ── Custom commands ─────────────────────────────────────────────────────
\newcommand{\R}{\mathbb{R}}
\newcommand{\E}{\mathbb{E}}
\newcommand{\norm}[1]{\left\|#1\right\|}
\newcommand{\KL}[2]{D_{\mathrm{KL}}\left(#1\,\|\,#2\right)}

% ── Theorem environments ────────────────────────────────────────────────
\newtheorem{theorem}{Theorem}
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{proposition}[theorem]{Proposition}
\newtheorem{corollary}[theorem]{Corollary}
\theoremstyle{definition}
\newtheorem{definition}{Definition}
\theoremstyle{remark}
\newtheorem{remark}{Remark}

\begin{document}
\maketitle

% ── Abstract ────────────────────────────────────────────────────────────
\begin{abstract}
We address [the problem] in the setting of [context].
Our main contribution is [method/result], which achieves [metric] on [benchmark],
outperforming prior work by [margin]. We provide theoretical guarantees showing [guarantee].
Code will be released at \texttt{https://github.com/anonymous/repo}.
\end{abstract}

% ── 1. Introduction ─────────────────────────────────────────────────────
\section{Introduction}
\label{sec:intro}

Start with the problem and why it matters.
\citet{vaswani2017attention} introduced the transformer architecture;
subsequent work \citep{devlin2019bert,brown2020language} showed...

\paragraph{Contributions.}
Our main contributions are:
\begin{enumerate}
  \item[(i)] A novel [method] that [property] (Section~\ref{sec:method}).
  \item[(ii)] Theoretical analysis proving \emph{[convergence rate]}
    (Theorem~\ref{thm:main}).
  \item[(iii)] Empirical evaluation showing \emph{state-of-the-art} on
    [benchmark] (Section~\ref{sec:experiments}).
\end{enumerate}

% ── 2. Related Work ─────────────────────────────────────────────────────
\section{Related Work}
\label{sec:related}

\paragraph{[Topic A].}
Prior work by \citet{lecun1998gradient} established [result].
More recently, \citet{he2016deep} showed [finding].

\paragraph{[Topic B].}
Our work differs from \citet{goodfellow2016deep} in that [key difference].

% ── 3. Method ───────────────────────────────────────────────────────────
\section{Method}
\label{sec:method}

\subsection{Problem Setup}

Let $\mathcal{X} \subset \R^d$ denote the input space and
$\mathcal{Y}$ the label space.
We observe $n$ i.i.d.~samples $(\mathbf{x}_i, y_i)_{i=1}^n$ drawn from
distribution $\mathcal{D}$ over $\mathcal{X} \times \mathcal{Y}$.
Our goal is to minimise the expected risk:
\[
  \mathcal{R}(f) = \E_{(\mathbf{x},y) \sim \mathcal{D}}[\ell(f(\mathbf{x}), y)]
\]

\subsection{Proposed Algorithm}

\begin{algorithm}[h]
  \caption{[Your Algorithm Name]}
  \label{alg:main}
  \begin{algorithmic}[1]
    \STATE \textbf{Input:} Dataset $\{(\mathbf{x}_i, y_i)\}_{i=1}^n$,
      learning rate $\eta > 0$, iterations $T$
    \STATE Initialise parameters $\theta_0 \in \R^p$
    \FOR{$t = 1, \ldots, T$}
      \STATE Sample mini-batch $\mathcal{B} \subset [n]$, $|\mathcal{B}| = b$
      \STATE Compute gradient $\mathbf{g}_t = \frac{1}{b}\sum_{i \in \mathcal{B}}
        \nabla_\theta \ell(f_\theta(\mathbf{x}_i), y_i)$
      \STATE Update $\theta_{t+1} = \theta_t - \eta \mathbf{g}_t$
    \ENDFOR
    \RETURN $\theta_T$
  \end{algorithmic}
\end{algorithm}

\subsection{Theoretical Analysis}

\begin{theorem}
  \label{thm:main}
  Under Assumptions [A1--A3], Algorithm~\ref{alg:main} satisfies:
  \[
    \E[\mathcal{R}(f_{\theta_T})] - \mathcal{R}^* \leq
    \frac{\|\theta_0 - \theta^*\|^2}{2\eta T} + \frac{\eta L \sigma^2}{2}
  \]
  where $L$ is the Lipschitz constant and $\sigma^2$ is the gradient variance.
\end{theorem}
\begin{proof}
  By the descent lemma for $L$-smooth functions and telescoping. $\square$
\end{proof}

% ── 4. Experiments ──────────────────────────────────────────────────────
\section{Experiments}
\label{sec:experiments}

\subsection{Setup}

\textbf{Datasets.}
We evaluate on [Dataset A] ($n$ = X, $d$ = Y) and [Dataset B] ($n$ = X).

\textbf{Baselines.}
We compare against: (1) Method A \citep{ref1}; (2) Method B \citep{ref2};
(3) Method C (our ablation).

\textbf{Implementation.}
All models trained on [hardware] with [framework]. Hyper-parameters selected via
cross-validation; see Appendix~\ref{app:hparams} for details.

\subsection{Main Results}

Table~\ref{tab:main} shows that our method achieves state-of-the-art on all benchmarks.

\begin{table}[t]
  \centering
  \caption{Comparison on [task]. Higher is better ($\uparrow$). Best in \textbf{bold}.}
  \label{tab:main}
  \begin{tabular}{lcccc}
    \toprule
    Method & Acc.~(\%) $\uparrow$ & F1 $\uparrow$ & Params & FLOPs \\
    \midrule
    Baseline~\citep{ref1} & 88.4 & 0.879 & 25M & 4.1G \\
    Method B~\citep{ref2} & 91.2 & 0.908 & 67M & 11.3G \\
    \midrule
    \textbf{Ours} & \textbf{94.7} & \textbf{0.943} & \textbf{18M} & \textbf{2.8G} \\
    \bottomrule
  \end{tabular}
\end{table}

\subsection{Ablation Study}

We isolate the contribution of each component in Table~\ref{tab:ablation}.

\begin{table}[t]
  \centering
  \caption{Ablation on [Dataset A].}
  \label{tab:ablation}
  \begin{tabular}{lcc}
    \toprule
    Variant & Acc. & $\Delta$ \\
    \midrule
    Full model & 94.7\% & --- \\
    w/o component A & 92.1\% & $-$2.6 \\
    w/o component B & 90.8\% & $-$3.9 \\
    \bottomrule
  \end{tabular}
\end{table}

% ── 5. Conclusion ───────────────────────────────────────────────────────
\section{Conclusion}
We proposed [method], achieving [result]. Future work includes [direction].

% ── References ──────────────────────────────────────────────────────────
\begin{thebibliography}{9}
\bibitem[Vaswani et al.(2017)]{vaswani2017attention}
Vaswani, A., Shazeer, N., Parmar, N., et al. (2017).
\textit{Attention Is All You Need}.
NeurIPS.
\bibitem[Brown et al.(2020)]{brown2020language}
Brown, T., Mann, B., Ryder, N., et al. (2020).
\textit{Language Models are Few-Shot Learners}.
NeurIPS.
\bibitem[He et al.(2016)]{he2016deep}
He, K., Zhang, X., Ren, S., Sun, J. (2016).
\textit{Deep Residual Learning for Image Recognition}.
CVPR.
\end{thebibliography}

% ── Appendix ────────────────────────────────────────────────────────────
\appendix
\section{Proofs}
\label{app:proofs}

\textbf{Proof of Theorem~\ref{thm:main}.}
Full derivation here.

\section{Hyper-parameter Settings}
\label{app:hparams}

\begin{table}[h]
  \centering
  \begin{tabular}{lcc}
    \toprule
    Hyper-parameter & Dataset A & Dataset B \\
    \midrule
    Learning rate & $10^{-3}$ & $5\times10^{-4}$ \\
    Batch size & 256 & 128 \\
    Epochs & 200 & 100 \\
    \bottomrule
  \end{tabular}
\end{table}

\end{document}

Is the ML Conference Paper (NeurIPS / ICML / ICLR style) template free?

Yes — every latexci template is free, open-source, and needs no signup. Open it in the browser editor or copy the full source.

How do I use this template?

Click “Open in editor” to load it straight into the live LaTeX preview, or copy the source into Overleaf or your local TeX editor (TeX Live, MiKTeX).