Read a paper without getting lost

You do not have to read a paper from the first word to the last. Your first job is to discover what kind of paper it is and whether it deserves deeper attention.

Use a seven-pass workflow

1. Title and abstract: What problem, method, and headline result does the paper claim?
2. Figures and captions: What was actually measured or built? Can you identify the main comparison?
3. Introduction: Why do the authors say the problem matters? What gap do they claim?
4. Main result: Find the table, figure, proof, or qualitative evidence that supports the central claim.
5. Conclusion and limitations: What do the authors believe follows—and where do they admit uncertainty?
6. Methods when needed: Read the parts required to judge or reproduce the evidence.
7. Return for detail: Only now resolve equations, unfamiliar procedures, appendices, or implementation choices relevant to your purpose.

At the end of the first five passes, decide whether to stop, save it as background, or study it closely.

This is a Matrix reading workflow, not a rule from a journal. Skip or repeat a pass when your purpose requires it.

Separate vocabulary trouble from contribution trouble

Unknown vocabulary is local. Highlight the term, infer its role from context, then consult a textbook, review paper, or reliable reference. You may still understand the paper’s contribution.

Contribution trouble is different: you cannot state what changed compared with prior work, what evidence supports it, or why the result matters. Return to the introduction, main figure, baseline, and conclusion. If those still do not connect, the paper may be unclear—or you may need a review paper first.

Read results skeptically

For every main result, ask:

• What is the baseline or comparison?
• What data, sample, or cases were included—and excluded?
• Does the metric represent the real goal?
• How much variability or uncertainty is visible?
• Is the improvement large in practical terms?
• Does the conclusion go beyond the tested population or conditions?

Do not assume “statistically significant,” “state of the art,” or a visually dramatic plot means the result is useful or broadly generalizable.

Use citations as a map

References in the introduction lead backward to foundational methods and competing explanations. “Cited by” links lead forward to replications, corrections, extensions, and later uses. Review papers group these routes into a field-level map.

Keep a small literature table: citation, question, method, dataset/sample, result, limitation, and relevance to your idea. Patterns across rows are more useful than a pile of PDFs.

Know when you are done

Your purpose determines depth. To decide whether a topic is promising, the main question and limitations may be enough. To reproduce the work, you need methods, parameter choices, data processing, and supplementary material. To cite a claim, read the original context—never rely on another paper’s one-sentence summary.