paper-writer

name	paper-writer
description	Write academic papers from experiment results using LaTeX. Use when experiments are complete and REPORT.md exists, when asked to write a paper, or when generating publication-ready documents in NeurIPS style.

Paper Writer

Guide for writing academic papers from experiment results using a two-stage process.

When to Use

• After experiments are complete and REPORT.md exists
• When explicitly asked to write a paper
• When generating publication-ready documents

Before Writing - Required Steps

IMPORTANT: Before writing any content, you MUST complete these steps:

1. Read the style guide: Review templates/paper_writing/lab_style_guide.md for comprehensive formatting and language conventions

2. Study example papers: Browse paper_examples/ to understand our lab's style:

   • Language patterns: Look at sections/1.introduction.tex or sections/introduction.tex
   • Table formatting: Look at tables/*.tex files
   • Figure layouts: Look at figures/*.tex files
   • Macro usage: Look at commands/*.tex files

3. Verify command templates: Command templates are pre-copied to paper_draft/commands/:

   • math.tex - Math notation macros
   • general.tex - Formatting macros (\para{}, colors, etc.)
   • macros.tex - Template for project-specific terms (customize for your paper)

4. CRITICAL: Reference example papers for FORMATTING and LANGUAGE STYLE only

   • Do NOT copy content, phrasing, or narrative structure
   • The example papers are in different research domains
   • Focus only on HOW things are formatted, not WHAT is written

5. Set paper author: Read .neurico/idea.yaml to find idea.metadata.author:

   • If metadata.author exists: use <author name> and NeuriCo
   • If no metadata.author: use NeuriCo

Two-Stage Writing Process

Stage 1: Outline Development

Before writing prose, create a detailed outline:

1. Skeleton: Section headers and subsection structure
2. Key points: Bullet points for each section (3-5 per section)
3. Evidence mapping: Link each claim to supporting data/figures
4. Citation placeholders: Note where references are needed
5. Figure/table planning: List required visuals

Save outline to paper/OUTLINE.md for review before proceeding.

Stage 2: Prose Writing

Convert outline to full prose:

1. Write section by section (don't jump around)
2. Expand each bullet into 2-4 sentences
3. Add transitions between paragraphs
4. Insert citations as you write
5. Create figures/tables as needed

Paper Structure (IMRAD Format)

1. Title

• Clear, specific, informative
• Conveys main finding or contribution
• No acronyms unless universally known

2. Abstract (150-250 words)

Follow this structure:

• Context/Problem (1-2 sentences): Why does this matter?
• Gap/Challenge (1 sentence): What's missing?
• Our approach (1-2 sentences): What did we do?
• Key results (2-3 sentences): What did we find?
• Significance (1 sentence): Why does it matter?

3. Introduction

Structure:

1. Hook (1 paragraph): Why does this problem matter?
2. Background (1-2 paragraphs): What do readers need to know?
3. Gap (1 paragraph): What's missing in existing work?
4. Contribution (1 paragraph): What do we provide? Be specific with bullets:
   • Contribution 1
   • Contribution 2
   • Contribution 3
5. Organization (optional): Brief roadmap of paper

4. Related Work

Organization strategies:

• By theme: Group papers by approach/concept
• By timeline: Historical development (less preferred)
• By relationship: How papers relate to ours

For each group:

• Summarize the approach
• Identify limitations
• Position our work: "Unlike X, we..." or "Building on X, we..."

5. Method/Approach

Essential elements:

• Problem formulation (formal if appropriate)
• Method description (clear enough to reproduce)
• Design justifications (why this choice?)
• Algorithm/pseudocode (if complex)
• Complexity analysis (if relevant)

6. Experiments

Structure:

1. Setup

   • Datasets: source, size, preprocessing
   • Baselines: what and why
   • Metrics: what and why
   • Implementation: hardware, hyperparameters

2. Main Results

   • Tables with clear captions
   • Statistical significance (confidence intervals or p-values)
   • Bold best results

3. Analysis

   • What do the numbers mean?
   • Why does our method work?
   • Where does it fail?

4. Ablations

   • Component contributions
   • Sensitivity analysis
   • Design choice validation

7. Discussion

Cover:

• Limitations (be honest and specific)
• Broader implications
• Failure cases and edge cases
• Connections to theory (if applicable)

8. Conclusion

Format:

• Summary (1 paragraph): What did we do and find?
• Key takeaway (1 sentence): What should readers remember?
• Future work (2-3 sentences): What comes next?

LaTeX Template

Style files (.sty, .bst) are copied to the paper_draft/ directory. The exact preamble (package name, options, bibliography style) is specified in your prompt - follow it exactly.

Template Structure

\documentclass{article}
% Conference style package - USE THE EXACT LINE FROM YOUR PROMPT
\usepackage{<style_package>}  % e.g., neurips_2025, icml2026, etc.

% Required packages - ALWAYS include these
\usepackage[hidelinks]{hyperref}  % Clickable links (REQUIRED)
\usepackage{booktabs}  % Better tables (REQUIRED)
\usepackage{graphicx}  % Figures
\usepackage{amsmath,amssymb}  % Math

% Import command files
\input{commands/math}
\input{commands/general}
\input{commands/macros}

\title{Clear Title That Conveys Main Contribution}

% Set author based on .neurico/idea.yaml metadata.author:
%   If metadata.author exists: \author{<author name> and NeuriCo}
%   If no metadata.author:     \author{NeuriCo}
\author{NeuriCo}

\begin{document}
\maketitle

\begin{abstract}
Your abstract here (150-250 words).
\end{abstract}

\section{Introduction}
...

\bibliography{references}
\bibliographystyle{<bib_style>}  % Use the style from your prompt

\end{document}

Table Formatting

\begin{table}[h]
\centering
\caption{Results comparing methods on [benchmark].
         Higher is better for all metrics.
         Best results in \textbf{bold}.}
\begin{tabular}{lcc}
\toprule
Method & Accuracy (\%) & F1 (\%) \\
\midrule
Baseline 1 & 75.2 {\scriptsize $\pm$ 0.3} & 72.1 {\scriptsize $\pm$ 0.4} \\
Baseline 2 & 78.4 {\scriptsize $\pm$ 0.2} & 75.8 {\scriptsize $\pm$ 0.3} \\
\midrule
Ours & \textbf{82.1} {\scriptsize $\pm$ 0.2} & \textbf{79.4} {\scriptsize $\pm$ 0.3} \\
\bottomrule
\end{tabular}
\label{tab:main_results}
\end{table}

Figure Formatting

\begin{figure}[h]
\centering
\includegraphics[width=0.8\linewidth]{figures/main_result.pdf}
\caption{Caption should be self-contained. Explain what is shown,
         highlight key observations, and note any important details.}
\label{fig:main_result}
\end{figure}

Citation Guidelines

BibTeX Format

@inproceedings{author2024title,
  title={Full Paper Title},
  author={Last, First and Last2, First2},
  booktitle={Conference Name},
  year={2024}
}

Citation Style

• Use \cite{key} for parenthetical: "...as shown previously (Author et al., 2024)"
• Use \citet{key} for textual: "Author et al. (2024) showed that..."

Lab Style Conventions

Quick Reference

These are the key conventions from our lab's writing style. See templates/paper_writing/lab_style_guide.md for complete documentation.

Language:

• Active voice: "We propose", "We examine", "We focus on"
• Clear and simple - prefer plain language over jargon
• Bold questions as organizers: {\bf what is hypothesis generation?}
• Specific quantitative claims: "8.97% over baselines"

Structure:

• Modular commands/ directory with math.tex, general.tex, macros.tex
• Import with \input{commands/math} etc.

Macros:

• Vectors: \va, \vb, ..., \vz (bold lowercase)
• Matrices: \mA, \mB, ..., \mZ (bold uppercase)
• References: \figref{}, \Figref{}, \secref{} (not raw \ref{})
• Method names: \newcommand{\methodname}{\textsc{Name}\xspace}

Tables:

• Always use booktabs (\toprule, \midrule, \bottomrule)
• Use \resizebox{\textwidth}{!}{...} for wide tables
• Use @{} at edges, \cmidrule(lr){x-y} for sub-headers
• Bold best results

Hyperlinks (Required):

• Always use \usepackage[hidelinks]{hyperref}
• All citations, refs must be clickable

Figures:

• Use 0.32\textwidth for 3-column subfigures
• Use \input{figures/legend} for shared legends
• Self-contained captions

Contribution Lists:

\begin{itemize}[leftmargin=*,itemsep=0pt,topsep=0pt]
    \item We propose...
    \item We conduct...
\end{itemize}

Output

Save to paper_draft/ directory with this structure:

paper_draft/
├── main.tex              # Main document
├── references.bib        # BibTeX citations
├── commands/
│   ├── math.tex          # Math notation macros
│   ├── general.tex       # Formatting macros
│   └── macros.tex        # Project-specific terms
├── sections/
│   ├── abstract.tex
│   ├── introduction.tex
│   └── ...
├── figures/              # Figure files (PDF preferred)
└── tables/               # Complex standalone tables

Compile with:

cd paper_draft && pdflatex main && bibtex main && pdflatex main && pdflatex main

Quality Checklist

Content

• Title reflects main contribution
• Abstract is self-contained (no citations, no undefined terms)
• Contributions clearly stated in introduction
• All claims supported by evidence
• Limitations honestly discussed
• Related work positions paper clearly

Technical

• Method reproducible from description
• All experimental details provided
• Statistical significance reported
• Ablations validate design choices

Presentation

• Figures have informative captions
• Tables are properly formatted
• All citations present and correct
• No placeholder text
• Consistent notation throughout
• Proofread for typos

Ethics

• Broader impact considered
• Potential negative uses discussed
• Data/model limitations noted

References

See references/ folder for:

• writing_guidelines.md: Section-specific writing advice

See assets/ folder for:

• paper_outline_template.md: Template for Stage 1 outline