Planning paper writing

Devi Parikh
4 min readAug 26, 2019

--

Better papers with less drama.

I have a strong aversion to avoidable drama and stress in life. I have developed a few systems that help me minimize said drama and stress.

One of them is for planning writing papers for conference deadlines.

As an added bonus, it also results in stronger papers on average.

I have addressed this post to students, and assumed that they are iterating on the paper with their advisor. But the ideas apply more generally to any collaboration. For more collaborators, I recommend agreeing on an order in which collaborators will take passes.

Principle 1: Iterate on the paper in a hierarchical fashion (coarse to fine).

Do not start with a full draft of the paper. There are two advantages to this:

  1. Your advisor may disagree with you on the nuances of the story or structure of the paper. It is better to agree on the higher order bits first to avoid having to redo large chunks of writing.
  2. It is easier to get started on a skeleton of the paper than on 8 pages worth of beautiful sentences and figures. Writer’s block is unlikely.

Principle 2: Iterate on small chunks at a time.

  1. It is less overwhelming for you to receive and for your advisor to give feedback.
  2. It allows for parallel progress. While your advisor is taking a pass on one section, you could be working on a draft for another.

Principle 3: Plan for multiple iterations on every section.

Principle 4: Schedule each iteration.

Let your advisor know when you’ll need them to take a pass on a section. That way they can plan ahead, and time spent waiting on your advisor’s pass is minimized.

Incorporate their constraints when scheduling iterations. E.g., what days of the week do they teach and are less likely to have time? Are they traveling?

Think about how long it might take to revise a section based on feedback. Incorporate your constraints. E.g., when do you typically have assignments due? Do you have exams or travel coming up before the deadline?

Caveat: This will work better if you and your advisor have good time management. This might help with that.

Methodology
Iterate on the paper in the following order.

  1. Paper title. May need to be revisited.
  2. Paper skeleton: Section and subsection headings.
  3. Intro skeleton: A bulleted list describing the point you want to make in each paragraph of the intro. No full sentences.
  4. Intro filled in: Fill it in with sentences. You now have the first draft of your paper’s intro. May need to iterate on this a couple of times.
  5. Related work skeleton: List of topics related to your work, a list of papers in each you plan to discuss, and a phrase stating what you plan to say about each.
  6. Related work filled in.
  7. Approach skeleton: Subsections and bulleted list in each charting out how you’ll walk the reader through your approach. Placeholders for figures.
  8. Approach filled in: Approach sections are often longer and denser than intro and related work. I recommend splitting filling in the approach into two chunks.
  9. Results skeleton: Datasets you’ll show results on, points covering the experimental setup, subsections with associated claims you will support, placeholder figures to illustrate expected trends, placeholder tables that list existing work/baselines/ablations you will compare to, list of analyses. No actual results, data, or sentences. Iterations on this may reveal additional required experiments.
  10. Results filled in: Actual sentences, data, results.
  11. Conclusion, abstract.

Do not move on to subsequent stages of a section till you and your advisor have converged on earlier stages. There is flexibility across sections. E.g., you can work on an outline of related work while your advisor is taking a pass on the filled in intro.

Recall, each stage may take multiple iterations till you converge.

Tool

I use an excel sheet like this to schedule and keep track of the iterations.

Sheet to schedule and track iterations on paper sections.

It lists the date on which you will send your advisor the section by 9:00 am (say). Your advisor will work on it on that date and get back to you by the end of the day.

Notice that the sheet spans ~4 weeks! So yes, start writing the paper at least 4 weeks before the deadline. The actual duration will vary depending on how many iterations go into each stage and how available your advisor is.

Disclaimers

  • My experience is primarily with AI conferences.
  • You may need to adapt this to your working style, the order in which you like to write the sections, the typical paper format of the conferences that you submit to, the topic of your specific paper, etc.
  • I have had limited success in getting students to follow this. To quite an extent, that is because I feel the need to give them the freedom to find their own working style. I was fortunate enough to have that freedom when I was a grad student. So I have not insisted on this as much as I could. Having said that, whenever students have followed this, outcomes, in my opinion, have been better on average. So perhaps I should be insisting on this more.

--

--

Devi Parikh

Research Director at Facebook AI Research. Associate Professor at Georgia Tech. Generative Artist. Artificial Intelligence.