AI.Cruz - Learn AI Through Interactive Visualizations

RESEARCH

Writing Your First Research Paper

A research paper tells one story: here is a problem, here is our approach, here is the evidence it works. IMRaD structure makes that story navigable for any reader.

1The IMRaD Structure

Almost every ML paper follows IMRaD: Introduction, Methods, Results, and Discussion. This structure is not arbitrary — it maps directly to the scientific method. IMRaD gives readers a mental model before they start reading: they know exactly where to find the motivation, the technique, the evidence, and the caveats.

Counter-intuitive writing order: write in reverse. Start with figures and tables (your data drives everything). Then write Methods (what you did). Then Results (what you found). Then Introduction (why it matters). Write the Abstract last — it is a compression of the whole paper, so write it when the whole paper exists.

Introduction
What problem? Why does it matter? What has been tried? What is missing? What do we contribute? (bullet list). Paper roadmap ("Section 2 presents...").

Methods
How does your approach work? Enough detail that someone could reimplement it. Notation, model architecture, training procedure, datasets.

Results & Discussion
What did you measure? How does it compare to baselines? Why does it work (ablations)? What are the limitations? What comes next?

2Abstract and Introduction

The abstract is the most-read part of any paper. A bad abstract means the paper is skipped. A strong abstract compresses the entire story into 150-250 words and answers five questions: problem, gap, approach, main result, significance.

Abstract formula

Sentence 1-2: State the problem and why it matters
Sentence 3: Gap in prior work
Sentence 4-5: Our approach (what we do, not why it works)
Sentence 6: Key quantitative result
Sentence 7: Significance / broader impact

Introduction formula

Para 1: Hook — motivate the problem vividly
Para 2-3: Prior work and why it falls short
Para 4: Our approach at high level
Para 5: Contribution bullet list (3-5 items)
Para 6: Paper roadmap

3Methods, Results, and Figures

The methods section is a reproducibility contract: a competent reader should be able to rerun your experiment from this section alone (plus cited baselines). The results section reports everything — not just the wins. Hiding negative results is how irreproducible research propagates.

Methods checklist
Dataset (size, splits, source), model architecture, training hyperparameters (all of them), evaluation metrics, statistical procedure (seeds, significance test).

Results checklist
Main comparison table (mean ± std), ablation table (remove one component at a time), failure analysis, compute cost if relevant.

Figures over tables
Learning curves, attention maps, qualitative examples — show don't just tell. A figure a reader can scan in 10 seconds beats a table they study for 2 minutes.

Design figures for black-and-white printing and colour-blind readers. Use markers and line styles in addition to colour. Label every axis. Caption every figure completely — a figure should be interpretable without reading the main text.

4LaTeX, Citations, and Submission

LaTeX is the standard format for ML research. Every major venue (NeurIPS, ICML, ICLR, ACL) provides an official LaTeX template. BibTeX manages references. The workflow is: write in Overleaf or VS Code + LaTeX Workshop, post a preprint to arXiv before or after submission, then submit to the venue.

LaTeX tools
Overleaf: collaborative cloud editor, no install required. VS Code + LaTeX Workshop: local, fast, integrates with Git. Always use the official venue template from day one.

Citation managers
Zotero: free, exports to BibTeX, syncs across devices. Mendeley: PDF annotator + reference manager. Export all references as a single .bib file and keep it under version control.

arXiv preprint
Post to arXiv (cs.LG or cs.AI) before the conference deadline for visibility. Most venues explicitly allow preprints. An arXiv version establishes a timestamp and gets your work circulated during review.

BibTeX formatter and citation key generator

from dataclasses import dataclass
from typing import List, Optional
import re

@dataclass
class Citation:
    authors: List[str]
    title: str
    year: int
    venue: str
    venue_type: str  # "conference", "journal", "preprint"
    volume: Optional[str] = None
    pages: Optional[str] = None
    arxiv_id: Optional[str] = None

def make_bibtex_key(citation: Citation) -> str:
    """Generate a consistent BibTeX key: firstauthorYYYYfirstword."""
    first_author_last = citation.authors[0].split()[-1].lower()
    first_word = re.sub(r'[^a-z]', '', citation.title.split()[0].lower())
    return f"{first_author_last}{citation.year}{first_word}"

def format_bibtex(citation: Citation) -> str:
    key = make_bibtex_key(citation)
    author_str = " and ".join(citation.authors)
    entry_type = "inproceedings" if citation.venue_type == "conference" else (
        "article" if citation.venue_type == "journal" else "misc"
    )
    lines = [
        f"@{entry_type}" + "{" + key + ",",
        f'  author    = "{author_str}",',
        f'  title     = "{citation.title}",',
        f'  year      = {citation.year},',
    ]
    if citation.venue_type == "conference":
        lines.append(f'  booktitle = "{citation.venue}",')
    elif citation.venue_type == "journal":
        lines.append(f'  journal   = "{citation.venue}",')
        if citation.volume:
            lines.append(f'  volume    = "{citation.volume}",')
        if citation.pages:
            lines.append(f'  pages     = "{citation.pages}",')
    if citation.arxiv_id:
        lines.append(f'  note      = "arXiv:{citation.arxiv_id}",')
    lines.append("}")
    return "\n".join(lines)

references = [
    Citation(
        authors=["Ashish Vaswani", "Noam Shazeer", "Niki Parmar", "Jakob Uszkoreit"],
        title="Attention Is All You Need",
        year=2017,
        venue="Advances in Neural Information Processing Systems",
        venue_type="conference",
        arxiv_id="1706.03762",
    ),
    Citation(
        authors=["Jacob Devlin", "Ming-Wei Chang", "Kenton Lee", "Kristina Toutanova"],
        title="BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding",
        year=2019,
        venue="North American Chapter of the Association for Computational Linguistics",
        venue_type="conference",
        arxiv_id="1810.04805",
    ),
    Citation(
        authors=["Yann LeCun", "Leon Bottou", "Yoshua Bengio", "Patrick Haffner"],
        title="Gradient-Based Learning Applied to Document Recognition",
        year=1998,
        venue="Proceedings of the IEEE",
        venue_type="journal",
        volume="86",
        pages="2278--2324",
    ),
]

print("=== Generated BibTeX entries ===\n")
for ref in references:
    print(format_bibtex(ref))
    print()

print("\n=== In-text citation keys ===")
for ref in references:
    key = make_bibtex_key(ref)
    print(f"\\cite{{{key}}}  →  {ref.authors[0].split()[-1]} et al. {ref.year}")