How to Write a Good Research Paper

A practical guide for graduate students

Abhishek Gupta

Associate Professor, ECE | The Ohio State University
Co-Director, IITB-OSU Frontier Center
Founder, Ensemble Control Inc.

Goals of Writing a Research Paper

  • Report new, generalizable and reproducible knowledge
  • Inform your audience about what is going on in the field and what your contributions are
  • Promote your work and promote yourself — marketing and branding

Side goal

I want someone to read my paper 30 years later.

Elements of a Research Paper

Main body

  • Introduction
  • Problem Formulation
  • Technical Material
  • Conclusion and Future Work

Appendix

  • Proofs
  • Experimental process / details
  • Any other information for recreating the proofs / experiments

One message: Every element has a job. Know what each section must accomplish before you write a word of it.

Know Your Target Audience

Type of paper Audience size Who reads it
Extremely theoretical This is my audience ~100 people Substantial mathematical training: faculty, math grad students, some industry research labs
Implementation & experimental ~1,000 people Graduate students and industry professionals who do experimental work
Survey ~10,000 people Starting graduate students, interested undergrads, non-experts
News article ~1,000,000 people Non-technical people, CTOs, CEOs, etc.

One message: The audience determines the notation, the pace, and the level of detail — decide who you are writing for before you start.

Goals for Individual Sections

Goals for Introduction

Good introductions…

  • put the result into perspective — literature survey: A did this, B did that, but we are doing X
  • state our contributions: how are your settings / results different from others?
  • use bullets, tables, and figures for better visualization
  • accurately describe the result: “under restricted conditions, our results outperform current state of the art”
  • convey the organization of the paper

Bad introductions…

  • list twenty citations with no story — “[3] did A. [7] did B. [12] did C.”
  • leave the reader guessing what is actually new
  • over-claim: “we solve the problem” when only a special case is solved
  • bury the contribution on page 4
  • read like a diary of the research, not a map of the paper

One message: By the end of the introduction, the reader must know what is new, why it matters, and where in the paper to find it.

Convey the Importance of the Problem

If this problem is not solved…

  • X people will die
  • Y cost will be incurred
  • Z risks need to be accounted for
  • A class of people will have a problem (ridesharing passengers will have large waiting times)

If the problem is solved…

  • New revenue-rich capabilities will be unlocked
  • New research directions will be created
  • Existing long-standing open problems are addressed
  • Existing knowledge is generalized via this framework

Also be careful about who this problem is for

General users, telecom companies, data center operators, utility companies, etc.

Goals for Problem Formulation

Good problem formulations…

  • select a reasonable notation — not too different from what the target audience understands
  • pick notation so readers can easily follow the arguments
  • use different notation for constants, spaces, sets, operators, etc.
  • clearly state the mathematical problem being solved

Bad problem formulations…

  • invent exotic notation for standard objects — and redefine symbols mid-paper
  • use \(x\), \(X\), \(\mathcal{X}\), and \(\mathbf{x}\) for four unrelated things
  • scatter assumptions across the paper instead of stating them up front
  • never actually write down the optimization problem / theorem target

One message: The reader should be able to restate your problem precisely after reading this section once.

Goals for Conclusion

Good conclusions…

  • reinforce the point of view of the paper
  • qualify the solution approach — X did not work because of A reasons, so we adopted Y
  • state shortcomings honestly, citing references that study problems addressing them
  • identify opportunities to strengthen the contributions — relax assumption A, run experiments under low-lighting conditions

Bad conclusions…

  • copy-paste the abstract with the verbs changed to past tense
  • hide the limitations and hope reviewers don’t notice
  • promise vague future work — “we will extend this to more general settings”
  • introduce a brand-new claim that appears nowhere else in the paper

One message: Stating your own shortcomings builds credibility — a reviewer who finds them first assumes you didn’t know.

Simple Tricks to Make Your Manuscript Readable

Simple Rules for Composition

Strive for: precision · clarity · familiarity · forthrightness · conciseness · fluidity · rhythm

Good writers…

  • keep the logical structure linear or close to linear: A ⟹ B ⟹ C
  • provide the page number when citing books / long papers
  • assume the reader cannot recall anything written 2 pages prior — and gently remind them

Bad writers…

  • create suspense — the key idea is revealed only in Section V
  • make the reader jump: “combining (12), Lemma 3, and the argument in [17]…”
  • cite a 600-page book with no page number
  • assume the reader holds the whole paper in their head

One message: A paper is not a mystery novel. Tell the reader the ending first, then show them the path.

Split Ideas into Small, Easily Digestible Chunks

  • Multiple short lines are better than one long line
  • Multiple short paragraphs are better than a long paragraph
  • Multiple short results leading to a long result is better than a 20-page proof
  • Two short papers are better than a long paper

One message: The reader’s working memory is small. Package every idea to fit inside it.

Write the Paper Backward

Research vs. writing

Research happens forward in time, but the paper is written with information in hindsight — so you need to write the paper backward from how your own research was conducted.

The dead ends, the failed attempts, the order of discovery — none of that is the reader’s problem. Start from the result, and lay the shortest logical path to it.