How I actually read an ML paper

Most papers are not written to be read start to finish. I used to try anyway, get stuck on the third equation, and quietly give up. Here is the loop I use now. It is boring on purpose.

Three passes, not one

The first pass is the title, the abstract, the figures, and the conclusion. Nothing else. The goal is one sentence: what does this paper claim, and what did it beat. If I cannot write that sentence, I have no business reading the rest yet.

The second pass is the method section, slowly. I redraw the main diagram by hand and label every arrow with a shape. If I do not know the shape of a tensor, I do not understand the step. This is where most of the real work happens.

The third pass is the parts I skipped: proofs, ablations, the appendix. By now I know which details actually matter, so I only read the ones that change the claim.

Rebuild the smallest version

Reading is not understanding. After the second pass I write the smallest possible version of the idea in code, with fake data, on my laptop. No training run, no cluster. Just enough to watch the numbers move the way the paper says they should.

If it does not behave, one of two things is true: the paper left something out, or I did. Both are worth knowing, and both go in the notes.

Write the note before you forget

The last step is a short note in my own words: the one-sentence claim, the trick that makes it work, and the one thing I would still not be able to explain to someone else. That last line is the honest one. It tells me exactly what to read next.

That is the whole method. Learn, build, document, repeat. This post is the document step.