ICLR 2026 LaTeX template (iclr2026_conference)

ICLR papers use iclr2026_conference.sty from github.com/ICLR/Master-Template. Scissor LaTeX has a free look-alike for drafting in your browser, with no account: start from the ICLR template, or load \usepackage{iclr2026_conference,times} in an article; the paper is an anonymous submission until you add \iclrfinalcopy. The same .tex compiles with the official kit.

Open the ICLR template in the editor Free, in your browser, no account

The template

The ICLR template is a representation-learning paper about spectral regularisation, with three authors, notation macros, a chart with a shaded band, an ablation table in booktabs style and author-year citations. Open the ICLR template in the editor to see all of it. It starts as the anonymous submission.

Turning it on

The kit is a package on a plain article. The official sample loads it together with times on one line:

\documentclass{article}
\usepackage{iclr2026_conference,times}

The year is read from the package name, so iclr2027_conference works too and the header shows that year. The page is always Letter and one column with 1.5 inch side margins, set in TeX Gyre Termes (a free Times substitute, downloaded the first time).

Submission and camera-ready

The kit has no package options. It switches with one command in the preamble:

The camera-ready version with \iclrfinalcopy

LaTeXCopyOpen in editor (opens in a new tab)

\documentclass{article}
\usepackage{iclr2026_conference,times}
\usepackage{amsmath}

\title{Spectral Regularisation for Stable\\ Representation Learning}
\author{Erika Mustermann \& Hans Mustermann \\
Department of Computer Science\\
Example University \\
\texttt{\{erika,hans\}@example.org}
\And
John Doe \\
Institute of Example Studies \\
\texttt{john.doe@example.org}}

\iclrfinalcopy

\begin{document}
\maketitle

\begin{abstract}
We penalise only the singular values of each layer that exceed a target, which keeps
training stable without slowing it down.
\end{abstract}

\section{Introduction}
Large singular values make a network sensitive to small changes of its input.
For a weight matrix $W$ we penalise $\sum_i \max(0, \sigma_i(W) - \tau)^2$.

\subsection{Contributions}
We test the penalty on 44 layer configurations.
\end{document}

Problems tab: Layout in the style of the ICLR style (final mode). Not official. To compile elsewhere, add the ICLR style files from github.com/ICLR/Master-Template.

What the layout sets

• The title is in large small capitals, flush left; the authors follow flush left. Separate them with \And or \AND; Scissor LaTeX puts each author block under the previous one, where the kit sets blocks joined by \And side by side when they fit.
• "Abstract" and all headings are in small capitals ("1 Introduction", "1.1 Contributions").
• Page numbers are centred at the foot of every page.
• Captions read "Figure 1:" and "Table 1:".

\input{math_commands.tex}, which the official sample uses for its notation macros, works as a normal \input, provided the file is in your project. The template defines a few macros of its own in the preamble instead; see macros.

Bibliography

The template uses \bibliographystyle{iclr2026_conference}, which Scissor LaTeX sets as author-year citations with \citet and \citep. See bibliographies.

What differs from the ICLR style

To submit, get iclr2026_conference.sty, its .bst file and math_commands.tex from github.com/ICLR/Master-Template and put them next to your .tex file; it compiles unchanged. See conference look-alike templates for the other layouts.

Every example on this page is typeset by Scissor LaTeX when the site is built. Last updated 2026-11-06.