Paper Structure

Agent-written drafts are usually grammatical and usually hollow. They restate the topic instead of making an argument. This is a section-by- section account of what each part has to accomplish.

Abstract

Four sentences, in this order, is enough: the problem, why existing work does not settle it, what you did, what you found with a number.

The most common failure is an abstract that describes the topic rather than the finding. "We study the effect of X on Y" is not a result. "X reduces Y by 12 percent, but only when Z holds" is.

Introduction

The job is to make the reader believe the problem is real before you propose anything.

1. The context, in two or three sentences, not a history of the field
2. The specific gap, stated so a reader can disagree with it
3. Your approach in one sentence
4. Your contributions as claims, not activities

"We conduct extensive experiments" is an activity. "We show that method X fails whenever the input distribution shifts by more than k" is a claim. Write claims.

Related work

Organise by idea, not chronology, and never as a list of summaries. The useful unit is: this line of work does A, it assumes B, and B does not hold in our setting.

If a paper is close to yours, say so plainly and say what is different. Reviewers already know that paper. Failing to cite it reads as ignorance or concealment.

Method

Enough for someone to reimplement it. Define your notation before using it. State assumptions explicitly, including the uncomfortable ones.

If the method is simple, say it is simple. Dressing a simple idea in notation to look sophisticated is transparent and it annoys reviewers.

Experiments

Structure: setup, then research questions, then a result per question.

• state datasets, baselines, metrics and seeds before any numbers
• one table or figure should answer one question
• report variance, not just means
• every number traces to a file (see numbers-traceable)

Include the experiment that did not work. Its absence is more conspicuous than its presence.

Limitations

Write it yourself, honestly, and do it before a reviewer does it for you. A limitations section that names the real weakness buys credibility for everything else. One that lists only "we did not test on more datasets" is read as evasion.

Conclusion

Two or three sentences: what you found, what it means, what is next. Do not introduce new claims here, and do not restate the abstract word for word.

Global failure patterns to check for

• Restatement. The same sentence appearing in abstract, intro and conclusion with different word order.
• Unsupported comparatives. "Significantly better", "much faster", "substantially improves", with no number and no test.
• Orphan artifacts. A figure or table nothing refers to.
• Contribution inflation. Four bullets that are one contribution split four ways.
• Passive vagueness. "It was observed that improvements occurred." Say who did what and by how much.