Loading lesson…
AI Research Foundations / Rigor & Communication
Communicating findings and surviving review.
Reviewed by Yuvaraj
A paper is not a diary of what you did; it is an argument for a claim, engineered so a busy, skeptical reader can grasp the contribution, judge the evidence, and reproduce the work. Peer review is the adversarial process that stress-tests that argument before it enters the scientific record. This lesson covers how to structure a paper so its contribution is unmistakable, how to write claims you can actually defend, and how review works from both sides, including the ethics that make the whole system trustworthy.
Every section of a well-built paper does a specific job in service of the central claim. The conventional structure survives because it maps onto how a critical reader evaluates a contribution: what is the problem, what is new, does the evidence hold, and where does it break.
| Section | The job it does | The question it answers |
|---|---|---|
| Title and abstract | State the contribution in its most compressed form | What did you do, and why should I care? |
| Introduction | Motivate the problem and state the specific claim and contributions | What is the gap, and what exactly do you claim to close it? |
| Related work | Locate the paper in the literature; establish novelty fairly | How is this different from what already exists? |
| Method | Describe the approach precisely enough to reproduce | What, exactly, did you build or propose? |
| Experiments | Provide fair, controlled evidence for the claim | Do the baselines, ablations, and metrics support the claim? |
| Results and analysis | Interpret the evidence, including where it fails | What do the numbers mean, and what are the limits? |
Ask about this lesson, or about anything in AI. Answers cite the lessons they draw on.
Finished this lesson?
Mark it complete to earn XP, keep your streak, and schedule a review.
| Limitations and conclusion | State honestly where the claim does not hold; summarize | When should I not trust or use this? |
The introduction carries the most weight. A strong one states the contribution as an explicit, bounded list, "we contribute (1) a method for X, (2) an analysis showing Y, (3) an evaluation on Z", so a reviewer can check each item against the rest of the paper. If a reader finishes your introduction unsure what you are claiming, no amount of experimental polish will save the paper.
Write the claim first, then the paper
Before drafting, write a single sentence: "This paper shows that ___." Everything in the paper either supports that sentence or should be cut. The abstract is that sentence expanded; the experiments are that sentence tested; the limitations are that sentence bounded. Papers lose reviewers when they contain three half-supported claims instead of one fully-supported one.
The fastest way to lose a reviewer's trust is a claim broader than your evidence. Every sentence in the introduction and abstract is a promise the rest of the paper must keep. The discipline is to scope claims to exactly what you measured, no more, and no less.
Words like "always," "solves," "proves," and "human-level" are promissory notes reviewers will call in. Replace them with the specific scope of your evidence: which task, which data, which metric, which comparison. A scoped claim is not weaker, it is credible, and it survives review because there is no gap between what you assert and what you showed. The limitations section is the natural companion to this discipline: stating plainly where your method fails is not a confession of weakness but a demonstration that you understand your own contribution's boundaries, and it pre-empts the reviewer's objection by making it first.
Reproducibility is part of the writing
The method and experimental sections are where reproducibility lives or dies. Report the hyperparameters, the data splits, the number of seeds, the compute, and where the code and data are. A claim a reader cannot reproduce from your paper is, to them, not yet established, and reviewers increasingly treat missing reproducibility details as a substantive weakness, not a formatting nit.
Peer review is the mechanism by which a community decides what to admit into its shared record. Understanding the lifecycle helps you write for it, and, eventually, serve in it.
A reviewer is not asking "is this perfect?" but "does this advance the field, and is the evidence sound?" Concretely they probe: Is the contribution novel and clearly stated? Are the claims supported by fair baselines and appropriate ablations (the disciplines from earlier lessons)? Are the results significant, not single-run noise? Is the work reproducible from the description? Are the limitations stated honestly? A paper can be technically correct and still be rejected for an unclear or overclaimed contribution, soundness and significance are judged together.
When you review, your job is to help the field and the authors, not to display cleverness. A good review summarizes the paper fairly in your own words (proving you understood it), separates major concerns that affect the decision from minor issues, is specific enough that the authors can act on each point, and calibrates its confidence honestly. Critique the work, never the authors; "the evaluation lacks a tuned baseline" is useful, "the authors are careless" is not. And review as you would want to be reviewed, thoroughly, on time, and in good faith.
A rebuttal is a targeted response, not a re-argument of the whole paper. Address the reviewers' concrete objections in order of importance: run the requested experiment if you can, correct genuine factual errors in the reviews politely and with evidence, and concede the points that are fair while explaining how you will address them. Stay measured, a defensive or dismissive tone loses borderline decisions that a calm, evidence-backed clarification would have won.
The entire system rests on honesty, because reviewers cannot re-run every experiment. Three obligations are foundational. Authorship must reflect genuine intellectual contribution, listing someone who did not contribute, or omitting someone who did, is misconduct; gift and ghost authorship both corrupt the record. Citation must credit prior work fairly and accurately, neither burying a competitor you built on nor citing padding you never used; misrepresenting what a cited paper showed is a distinct wrong. Honest reporting means presenting results as they are, no cherry-picked seeds, no hidden failed runs, no metric chosen after the fact, no plagiarized text or figures. Confidentiality binds reviewers too: a paper under review is privileged, not to be shared or used for your own advantage.
Common mistakes