Write research papers to be *skimmed*

Don't write your paper for somebody to read it, write for somebody to skim it.

This is my version of a rule I've heard many times before in different forms.

The justification for this rule is that most papers are very niche, and few people that will care enough to read every little detail. However, a greater number of people will want to read the paper at a higher level to understand the key idea, and an even greater number will just read the title / abstract to get a "one sentence" version of the paper. Therefore, out of the total number of hours that readers will spend on your paper, most of those hours will be spent not fully reading the paper. This is your audience, and you should write to this audience.

My way of operationalizing this rule is to write for a reader that is skimming the paper. A reader skimming the paper would appreciate:

• A key figure explaining the method (if possible, this is often hard to do well).
• A short introduction which can be read as a standalone summary of the paper.
• Section headers that clearly summarize each section's content.
• Figure/Table captions that are fairly self-contained.
• Important bits of text in bold or italics to catch attention visually.

This is not the only way to operationalize this principle. Another version I've heard¹ is something like "spend equal time on the title, the abstract, and the main text". It also gets at the truth that most readers won't actually read the main text, and you should explicitly account for this audience when writing the paper.

However, I prefer my version because it feels more actionable. It's focused on the product, not the time spent (which is just a proxy for the state of the finished product). Titles and abstracts are short, and there is arguably only so much you can do (both good and bad) for choosing a good title/abstract. However, the main text of a paper is much longer, and it's really easy to default into writing a narrative that makes sense end to end but isn't easy to skim.

Ultimately, regardless of which version of this rule you subscribe to, the lesson is clear: write papers for an audience isn't actually going to read the paper.

PS1: have I followed my own advice?

I think the last 2 papers from my PhD are ok: not amazing, but not terrible.

• Tanimoto Random Features for Scalable Molecular Machine Learning has a short intro and fairly clear section headers. The main disadvantage is that the actual algorithm isn't described very clearly in the paper.
• Retro-fallback: retrosynthetic planning in an uncertain world is arguably a bit worse, even though I liked this paper better. The sections introduce a bunch of parts of the algorithm in different sections before combining them together in §4.2, but I don't think this narrative is easy to understand from skimming the section titles alone. That being said, we did colour code sections, and the intro is fairly succinct.

PS2: do LLMs change this calculus at all?

Arguably the existence of LLMs to summarize research papers on demand makes the exact way that papers are written less important. Instead of skimming papers, researchers can feed it into an LLM and ask for a summary. All things being equal, I think this makes my advice less important than it used to be. However, I still endorse it, for 2 reasons:

1. Putting a paper into an LLM has non-zero cost (time and money, since tokens aren't free). There will be some fraction of your audience that will want to skim the paper without asking an LLM (or will skim the paper to decide whether to give the paper to an LLM).
2. LLM summaries usually don't depart much from the framing of the original paper (especially for cheaper LLMs). Clearly framing the important points in a way that makes the paper skimmable will probably also result in better summaries.