How do you take notes from research papers without copying the whole paper?

1. Write the reading question before opening a blank notes page

“Summarize this paper” supplies no filter, so every sentence can feel important. State why you are reading it: “Is this method suitable for a low-latency product?” “Does this result support my design?” or “Which step must I reproduce?” The same paper should produce different notes for different jobs.

If you do not yet have a concrete question, triage instead of taking full notes. Keshav's first pass checks the title, abstract, introduction, headings, conclusion, and references so you can judge category, context, correctness, contribution, and clarity before investing further.

• Deciding whether to continue: keep one relevance sentence.
• Comparing approaches: keep the claim, baseline, metric, and boundary.
• Reproducing the work: add data, equations, hyperparameters, and implementation details.
• Citing the work: save the source location and the exact scope it supports.

2. Map the paper on the first pass; do not transcribe it

Spend five to ten minutes identifying the problem, method entry point, decisive figure or table, and conclusion. Temporary marks inside the PDF are fine. Moving every paragraph into a separate document turns recording into a substitute for judgment and leaves you with a harder-to-search copy of the paper.

On the second pass, read only the sections that bear on your question. Keshav recommends inspecting figures, noting key points, and marking relevant references at this depth. If your task does not require proofs, appendices, or every experiment, completeness is not a reason to copy them.

3. Keep a four-line permanent card with evidence locations

The four lines are a filter, not the paper's official structure. Claim asks what the authors want you to believe. Evidence asks which result most directly supports it. Limitation asks where the conclusion should stop. Work implication asks which decision changes for you.

Limit each line to one or two sentences. Add a page, section, table, or equation to the claim and evidence. Separate author-stated limitations from your inference, and label the work implication as your judgment. The card handles navigation and decisions; the PDF retains detail.

• Claim: one sentence, without blending background, method, and result.
• Strongest evidence: one decisive experiment or derivation with a location.
• Biggest limitation: prefer an author-stated boundary; label inference.
• Work implication: name the choice, test, or next step that changes.

4. Worked example: compressing the Transformer paper

Reading question: Can an attention-only architecture compete on machine translation without recurrence while improving training parallelism?

Claim: the Transformer models sequences using attention rather than recurrence or convolution. Strongest evidence: the abstract and translation results report 28.4 BLEU on WMT 2014 English–German and compare training cost with prior systems. The durable note should point to the abstract and result table, not reproduce the full table.

Biggest limitation: the paper primarily validates the architecture on the translation tasks and configurations available at the time; it does not prove superiority for every sequence task, context length, or deployment environment. Work implication: if training parallelism is your constraint, continue into the architecture and complexity analysis; reproduce only if you need the equations, layer counts, optimizer, and training setup.

Leave in the PDF: every tensor shape, full hyperparameter list, every baseline score, and related work that does not change the current decision. The information remains available without occupying the decision card.

5. Short notes are not proof of learning

Slotte and Lonka support the direction that summary-style notes can outperform text-order or verbatim notes, but their participants were high-school graduates reading one philosophical text. That does not validate this four-line card across every discipline and workplace. Dunlosky and colleagues also rated summarization low in general utility: skilled summarizers may benefit, while untrained learners do not benefit consistently.

After completing the card, run a no-prompt recall check. Close the PDF and explain the claim, why the evidence supports it, and where the boundary lies. If you cannot, return to the evidence location instead of copying more sentences. For durable memory, turn the card into retrieval questions and revisit them over time.