ML Conference Manuscript Kit LaTeX Template

\documentclass[10pt]{article}
\usepackage[margin=1in]{geometry}
\usepackage[T1]{fontenc}
\usepackage[utf8]{inputenc}
\usepackage{lmodern}
\usepackage{amsmath,amssymb}
\usepackage{booktabs}
\usepackage{hyperref}

\title{<<paper_title>>}
\author{<<author_name>>\\<<institution_name>>\\\texttt{<<contact_email>>}}
\date{}

\begin{document}
\maketitle

\begin{abstract}
We present a compact conference-paper scaffold for machine learning manuscripts with emphasis on reproducibility reporting, ablation structure, and concise experimental claims.
\end{abstract}

\section{Introduction}
Recent deployment settings reveal strong train-test mismatch. This paper studies distillation methods that remain stable when data drift and noisy labels co-occur.

\section{Method}
Given teacher logits $z_t$ and student logits $z_s$, we optimize
\[
\mathcal{L}=\alpha\,\mathcal{L}_{\mathrm{CE}}+(1-\alpha)T^2\,\mathcal{L}_{\mathrm{KL}}(\sigma(z_t/T),\sigma(z_s/T)).
\]
We further add confidence gating for uncertain pseudo-labels.

\section{Experiments}
\subsection{Setup}
Datasets include two synthetic shifts and one real-world benchmark. Metrics are top-1 accuracy and expected calibration error.

\subsection{Results}
\begin{table}[h]
\centering
\begin{tabular}{@{}lcc@{}}
\toprule
Model & Accuracy & ECE \\
\midrule
Teacher baseline & 81.7 & 0.082 \\
Vanilla distillation & 83.1 & 0.071 \\
Proposed method & 85.4 & 0.049 \\
\bottomrule
\end{tabular}
\caption{Illustrative performance summary.}
\end{table}

\section{Discussion}
Observed gains come primarily from lower-confidence regions, suggesting that gating mitigates harmful teacher overconfidence.

\section{Conclusion}
The scaffold supports clear conference submissions and can be adapted to different review templates while preserving technical structure.

\section*{References}
\begin{enumerate}
  \item Example reference A.
  \item Example reference B.
\end{enumerate}

\end{document}