How to Structure a Research Paper

Readers of a research paper do not read it like a novel. A reviewer skims the title, the abstract, the last paragraph of the introduction, the results table, and the captions, then decides whether the method is worth their time. Structure is how you put the answers in those places. A beautiful model description that never states what was compared will be read as unfinished, because for this reader it is.

The title and abstract carry the comparison

The title should name the problem and, when you can do it without hype, the kind of result. The abstract is one paragraph with four beats: what is wrong or missing, what you did, what happened on the stated comparison, and what that does not mean. A number belongs in the abstract only if the results section contains it, on the split you will defend. “We propose a novel framework and achieve state-of-the-art results” is not an abstract. It gives the reviewer nothing to look up in the tables.

The introduction ends in contributions, not in a roadmap alone

One page is enough for most papers. Open with the problem a neighbouring specialist recognises. Say what the closest work already does. Then state the hole in a sentence a reader could check against your related-work section. End with two or three contribution bullets. Each bullet should be a result or a specific artefact, not “we explore” and not “we provide insights”. A final sentence can point at the section order. A paragraph that only says “the rest of this paper is organised as follows” is not a substitute for the bullets.

Related work is an argument, not a reading list

Group prior work by the approach, and end each group with what it does not establish that you need. The baselines you will report against have to appear here, with enough detail that the results table does not introduce a stranger. A list of papers in citation order, each given one polite sentence, does not position anything. If you have not done the search that justifies the hole, do that before you write this section. How to find a research gap is the same test.

Method, then the experiment that tests it

The method section describes what someone would reimplement: data in, model or procedure, training or study protocol, and what was held fixed. The experiment or study section describes the comparison: datasets or participants, splits, baselines, metrics, and the number of runs. Mixing them into one chronological lab diary makes it hard to see which details are the method and which are the test. For a machine-learning comparison, those experimental lines should already exist as the protocol in How to design a machine learning experiment.

Results report the comparison you precommitted to

Lead with the primary metric and the baseline. Put the spread next to the mean when runs vary. Ablations follow, each tied to a contribution bullet. Failures that change how a reader should use the result belong in the paper, not only in a footnote. A discussion then says what you will not claim: other datasets, other sites, a causal story you did not test, a deployment you did not run.

Fit the shape to the venue

A journal article can afford a fuller method and a longer related-work argument. A conference paper with a hard page limit has to move proofs, extra datasets, and implementation detail into an appendix and leave the comparison in the main paper. Choose the venue early enough that you are cutting toward its readers. How to choose a journal is that decision. An eight-page conference version is a different cut of the same comparison, not a different claim.