How to Read a Paper
A two-page editorial whose premise is economic: researchers read hundreds of hours of papers yearly, the skill is never taught, and the default strategy — front-to-back plowing — wastes most of that time. Keshav’s replacement is a budgeting discipline: three passes of increasing cost, with an explicit exit decision after each, so depth is spent only where it earns.
The three passes
Pass 1 — triage (5–10 min). Title, abstract, introduction, section headings, conclusions, a glance at references. Output is the five Cs: Category, Context, Correctness (do assumptions look valid?), Contributions, Clarity — and a decision: read on, or stop. Most papers stop here, rightly.
Pass 2 — comprehension (~1 hour). Read for content, skip proofs; mark unread references; scrutinize figures — mislabeled axes and missing error bars are his tell for “rushed, shoddy work.” Exit state: able to summarize the thrust with supporting evidence to someone else. Right depth for papers near, but not in, your specialty. Failing to understand here forks three ways: shelve it, return after background reading, or escalate.
Pass 3 — virtual re-implementation (1–5 hours). Recreate the work from the authors’ assumptions and diff your reconstruction against theirs. The diff is the instrument: it surfaces innovations, hidden assumptions, missing citations, and technique flaws — and transfers the authors’ proof and presentation tools into your repertoire. Exit state: reconstruct the paper’s structure from memory; required for reviewing.
The survey procedure
Bootstrap a literature survey by iterating the passes at collection scale: (1) search engines → 3–5 recent papers, one pass each, read their related-work sections; a survey paper found here ends the process. (2) Shared citations and repeated author names in the bibliographies identify key papers and people; where those people publish recently identifies the field’s top venues. (3) The venues’ recent proceedings supply the remaining high-quality work; two passes over the collection, iterating on any key paper they all cite that you missed.
The inversion for writers
The method’s sharpest corollary points at authors: reviewers and readers give a paper exactly one pass by default. A paper whose gist doesn’t survive pass 1 — coherent headings, a comprehensive abstract — gets rejected or, post-publication, simply never read. Writing for the pass structure is not gaming reviewers; it is the same economics from the other side.
Assessment
- Durable: the exit-decision structure — read-depth as an explicit resource allocation with checkpoints — and the re-implementation test for deep reading, which is method-independent.
- Era-bound: the survey mechanics (Google Scholar/CiteSeer keyword bootstrapping) date to 2007 and are now the most automatable step, though the shared-citation/key-people heuristic transfers intact to whatever does the searching.
- In this library: pass 1’s five Cs are what a catalog entry should let you answer without opening the blob, and pass 3’s “reconstruct from memory” is the honest bar for what a synthesis note ought to enable.