📄 How to Read Research Papers

“You don’t read papers — you interrogate them.”

Reading research papers is a learnable skill, not a talent. The average senior PhD student reads 3–5 papers per week and can extract the core contribution of any paper within 15–20 minutes. This guide teaches the systematic methods that experienced researchers use — from skimming to deep comprehension — and the tools to manage what you read.

1. Why Naive Reading Fails

Most graduate students read papers the wrong way: linearly, from abstract to conclusion, trying to understand every sentence before moving on. This approach has two critical flaws:

It is inefficient: You spend equal time on filler and on critical contributions.
It lacks context: Without a framework for where this paper fits in the literature, individual sentences are difficult to interpret.

The structured methods below fix both problems.

2. The Three-Pass Method

Developed by S. Keshav (How to Read a Paper), this method is the most widely taught in CS graduate programs.

Pass 1 — The Bird’s-Eye View (5–10 minutes)

Goal: Decide whether the paper deserves further reading.

Read only:

Title, abstract, introduction
Section headings and subheadings
Conclusion
References (skim for familiar names and venues)

Output: Answers to the five Cs:

Category: What type of paper is this? (Measurement, analysis, new system, survey?)
Context: What related work does it build on?
Correctness: Do the assumptions seem valid?
Contributions: What are the paper’s main claims?
Clarity: Is the paper well-written enough to understand?

Pass 2 — The Engaged Reading (30–60 minutes)

Goal: Understand the main content, not the implementation details.

Read carefully, but skip proofs and mathematical derivations for now:

Read the Introduction carefully — it contains the paper’s argument in compressed form.
Study all figures, tables, and graphs. Caption text is especially information-dense.
Annotate key claims, methods, and results.
Note any unfamiliar terms — look them up immediately.

Output: A 3–5 sentence summary of the paper in your own words. If you cannot produce this, you have not understood the paper.

Pass 3 — The Deep Dive (1–5 hours)

Goal: Virtually re-implement the paper.

Read every section, including proofs and appendices.
Challenge every assumption and experimental choice.
Identify the paper’s weaknesses.
Reproduce the key equations and algorithms on paper.

Output: Full understanding sufficient to (a) explain the paper to someone else, and (b) build on it in your own research.

[!TIP] Not every paper warrants Pass 3. Pass 1 for most papers, Pass 2 for relevant ones, Pass 3 only for papers directly in your research core.

3. Critical Reading Framework

Good researchers do not just understand papers — they critique them.

Questions to Ask During Reading

On the Problem

Is this problem important and well-motivated?
Is the problem statement precise, or is it vague enough to allow easy wins?
What would happen if this problem were left unsolved?

On the Method

Are the assumptions realistic?
Is there a simpler baseline that would achieve similar results?
What are the failure modes of this approach?

On the Evaluation

Is the evaluation on a benchmark that the authors may have overfit to?
Are the baselines fair and up-to-date?
Are the improvements statistically significant?
Do the experiments support the claims in the introduction?

On the Contribution

What is the actual novel contribution vs. the claimed contribution?
Could this result be an artifact of experimental setup?
How does this change what you thought you knew?

4. Annotation Strategies

Margin Annotation System

Use a consistent annotation vocabulary across all papers. Suggested symbols:

Symbol	Meaning
`★`	Key claim or contribution
`?`	Something I don’t understand yet
`!`	Surprising or counterintuitive result
`→`	Leads to a follow-up question
`✗`	I disagree with or doubt this
`📖`	New reference I should read
`💡`	My own idea sparked by this

Digital Annotation Tools

Zotero + PDF reader — Annotate PDFs and sync notes directly to your library. Annotations are searchable.
Hypothes.is — Web annotation layer for HTML papers (arXiv, ACM, IEEE).
Readwise Reader — Highlight and annotate PDFs and web articles with AI-assisted synthesis.
Obsidian — Write permanent notes linked to paper entries.

5. Building a Paper Notes System

Reading without note-taking is reading with a leaky bucket. A systematic notes system transforms consumed papers into reusable research capital.

The Atomic Note Template

For every paper you read at Pass 2 or deeper, create a note with this structure:

# [Paper Title] — [Year]

**Authors**: 
**Venue**: 
**Link**: 
**Read Date**: 

## Core Contribution (1–3 sentences)

## Key Method

## Key Results

## Limitations / Weaknesses

## My Questions

## How This Relates to My Research

## References to Follow Up

[!NOTE] The “How This Relates to My Research” field is the most important. If you cannot write something here, the paper may not belong in your permanent notes at all.

The Zettelkasten Approach

The Zettelkasten method (popularized by sociologist Niklas Luhmann, who wrote over 70 books using it) builds a network of atomic notes that link to each other. In Obsidian, this creates a visual knowledge graph where concepts in your research area are interconnected.

6. Managing Your Reading Queue

Literature Discovery Pipeline

New paper alert (Google Scholar, arXiv, Twitter)
           ↓
    Pass 1 filter (5 min)
           ↓
    Add to Zotero library
           ↓
    Tag and categorize
           ↓
    Schedule for Pass 2/3 based on relevance

Setting Up Paper Alerts

Google Scholar Alerts — Set keyword and author alerts to receive email notifications of new publications.
arXiv.org RSS feeds — Subscribe to specific subject areas (cs.LG, cs.CV, cs.NI, etc.) via RSS.
Semantic Scholar Research Feeds — Personalized paper recommendations based on your reading history.
Papers With Code — Tracks new papers with associated code and leaderboard results.
Twitter/X Lists — Follow the top 20 researchers in your area; they share the important papers as they come out.

7. Speed and Efficiency

Increasing Reading Speed

Read the abstract and conclusion first — they contain 80% of the paper’s message.
Read figures before text — figures are the paper’s argument compressed to its most visual form. If you understand all figures, you likely understand the paper.
Use 1.5× playback speed for video explanations — Many major papers now have Papers with Code explanations or YouTube walkthroughs.
Read in batches — Read 3–5 papers in the same cluster consecutively. Context from Paper 1 makes Paper 2 dramatically faster.

When to Stop Reading

If a paper is clearly outside your research scope after Pass 1 — stop.
If the methodology is so flawed that the results are meaningless — stop.
If a better, more recent paper has superseded it entirely — read the newer one instead.

8. Talking About Papers

Reading is passive. The test of understanding is explanation.

Weekly Lab Reading Groups

A structured weekly reading group where members take turns presenting a paper is one of the highest-leverage activities in a research lab. The presenter gains deeper understanding through the act of teaching; the audience gets a curated survey.

Paper Presentation Structure

A 20-minute paper presentation should cover:

Motivation (2 min): Why does this problem matter?
Prior art (2 min): What did people do before this paper?
Core contribution (5 min): What is genuinely new?
Key results (5 min): Walk through the most important figures and tables.
Critique (3 min): What are the weaknesses?
Discussion (3 min): How does this affect our own research?