Writing the Paper: Structure & Story
The anatomy of an AI paper, the order to write it in, an abstract formula, the four-paragraph introduction, contributions, related work, method, experiments, limitations, and sentences that are easy to read.
A paper is an argument
A paper is not a diary of what you did; it is an argument for one main claim, with evidence. Every section serves the claim. Most AI papers share a shape:
| Section | Job | Typical length (8-page paper) |
|---|---|---|
| Title | the claim or question, specific and searchable | < 15 words |
| Abstract | the whole paper in miniature | 150–250 words |
| Introduction | why it matters, what is missing, what you did, what you found | about 1 page |
| Related work | position against prior work: how you differ | 0.5–1 page |
| Method / setup | enough detail to reproduce | 1.5–2 pages |
| Experiments & results | evidence for each claim, with figures and tables | 2.5–3 pages |
| Analysis / discussion | why it works, when it fails | 0.5–1 page |
| Limitations, conclusion | honest boundaries; the takeaway | 0.5 page |
Many venues also require an ethics or broader-impact statement and a reproducibility checklist; appendices hold extra results and details.
Watch it: The shape of a paper: an hourglass
A short narrated animation. Press ▶ Play, or step through it with ⏮ ⏭. Tick 🔊 narrate to hear each step.
Interactive: this part runs in your browser.
Write in this order
- Figures and tables first: sketch the key figure and main results table. They are the evidence; the text explains them.
- Method and experimental setup: the easiest parts to write while fresh.
- Results and analysis: one paragraph per claim: state it, point to the evidence, interpret it.
- Introduction: now you know exactly what you found.
- Related work, limitations, conclusion.
- Abstract and title last.
Then revise in passes: structure (does each section do its job?), paragraphs (does each start with its point?), sentences (are they clear?), and finally typos and formatting. Ask someone outside the project to read it: what they misunderstand is what you must fix.
The abstract formula
A strong abstract answers five questions in about one sentence each:
- Context: what area and why it matters.
- Problem / gap: what is missing or wrong today.
- Approach: what you did (the method or study), concretely.
- Results: the main findings, with numbers and the comparison.
- Implication: why it matters; what readers should do differently. Mention code and data release.
Avoid vague claims ("significantly improves", "novel framework") without numbers. Avoid undefined acronyms. Every claim in the abstract must be supported in the paper.
Try it: build and check an abstract
Assemble an abstract from candidate sentences for each part, and see a reviewer-style check of what is missing or overclaimed.
Interactive: this part runs in your browser.
The introduction in four paragraphs
- The problem and why it matters: start broad enough to motivate, then narrow quickly. Concrete beats grand.
- The gap: what existing approaches do and exactly where they fall short (with citations). This sets up your contribution.
- What you do: your approach and the key idea, in plain words. Optionally a teaser figure.
- What you found, and the contributions: main results with numbers, then a short bulleted list of contributions ("We (1) …, (2) …, (3) release …"). Each contribution must be backed by a section.
Reviewers often decide their opinion after the introduction and Figure 1. Spend disproportionate time on them.
Related work, method, results, limitations
- Related work: organise by theme, not paper by paper; for each theme, say how your work differs. Be generous and accurate: reviewers are often the authors you cite.
- Method: define notation once, state assumptions, give the algorithm or equations, and every detail needed to reproduce (moved to the appendix if long).
- Results: each paragraph makes one claim, points to the table or figure, and interprets it. Include the variation (seeds, CIs) and the statistical test. Report results that do not favour you too.
- Analysis: ablations, error analysis, examples where it fails, sensitivity to hyperparameters.
- Limitations: settings not tested, assumptions, compute, data biases, potential harms. Honest limitations increase reviewers' trust; hiding them invites rejection when a reviewer finds them.
Sentences that are easy to read
- Put the point first: topic sentence, then support.
- One idea per sentence; prefer active voice and concrete verbs ("We train…", "X reduces errors by…").
- Use the same term for the same thing throughout; define it once.
- Cut filler ("it is important to note that", "in order to"). Replace "significantly" with a number unless you mean statistical significance.
- Guide the reader: "We first…, then…". Figures and tables must be understandable from their captions alone.
- Writing in your second language is normal in AI research. Simple, precise English is the goal; tools such as grammar checkers are fine, and many venues allow LLM help with language editing, but you remain responsible for every sentence (see Integrity).
Quiz
1. Which part should usually be written last?
2. Which abstract sentence is strongest?
3. Related work should be organised…
4. Why include a limitations section?
5. Each item in your contributions list should…
📚 Further reading
Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.
- S. Peyton Jones, How to Write a Great Research Paper ↗ talk and slides
- J. Eisner, Advice for Research Students ↗ a curated collection of advice on reading, writing and research