1Why Most Readers Waste Time
Reading a paper front-to-back like a novel is the single biggest mistake new researchers make. A 12-page ML paper can take 4-6 hours to fully absorb — but 80% of papers you encounter are not worth that investment. The first skill is not how to read; it is when to stop. The three-pass method, formalised by S. Keshav, solves this by decoupling paper triage from deep understanding.
The average working researcher reads 200-300 papers per year. At 5 hours each that is impossible. At 30 minutes average — using a tiered reading strategy — it becomes manageable. The key insight is that most papers can be triaged in under 10 minutes: title, abstract, introduction, figures, conclusion.
✗ Common mistakes
- ✗ Start at section 1, read to the end in one pass
- ✗ Skip figures and tables (they carry most information)
- ✗ No notes — rely on memory
- ✗ Read every paper to the same depth
✓ Better approaches
- ✓ Scan abstract and conclusion first to assess value
- ✓ Study figures before reading prose
- ✓ Take structured notes in your own words
- ✓ Match reading depth to paper relevance
2The Three-Pass Method
Keshav's three-pass method gives every paper exactly the depth it deserves. Most papers only get a first pass. Core papers in your area get all three.
Pass 1 — 15 to 30 min
Read title, abstract, introduction, section headings, and conclusion. Glance at math and figures but do not study them. Goal: understand the central claim and decide if the paper is relevant.
Pass 2 — 1 to 2 hours
Read carefully but skip proofs and derivations. Study every figure and table. Understand the approach and the experimental setup. Goal: be able to summarise the paper to a colleague.
Pass 3 — 4 to 5 hours
Virtual re-implementation. For every design choice, ask "why this and not that?" Challenge every assumption. Work through every proof. Goal: deep understanding of strengths and weaknesses.
Pass 1 deliverable: can you fill in these five Cs? Category (what type of paper?), Context (what papers does it relate to?), Correctness (do the assumptions seem valid?), Contributions (what are the main claims?), Clarity (is it well written?).
3Paper Structure: Where to Find What
ML papers follow a predictable structure. Knowing where to look saves enormous time. The abstract answers what. The introduction answers why. The methods section answers how. The experiments section answers does it work. The conclusion answers what next.
Abstract
4-8 sentences. Problem, approach, key result, significance. If the abstract is unclear, the paper is probably unclear too.
Introduction
Motivation and gap in prior work. The contribution list near the end is gold — authors explicitly state what they claim is new.
Related Work
What else exists in this area. Often skipped on first pass, but critical for placing the paper in context and finding other key references.
Methods / Model
The actual contribution. Most math lives here. On pass 1, read the section headings only. On pass 3, work through every equation.
Experiments
Baselines, datasets, metrics, ablations. The ablation table is often more informative than the main results table.
Conclusion & Appendix
Limitations and future work. Appendix contains proof details, extra experiments, and implementation specifics often omitted from the main text.
4Active Reading Techniques
Passive reading — eyes on page, mind elsewhere — produces zero retention. Active reading means writing while you read: annotate margins, summarise each section in one sentence, and question every claim. The goal is not to understand the paper once but to be able to recall its core ideas a month later.
Annotation tools
- ✓ Zotero — free reference manager with PDF annotations
- ✓ Notion / Obsidian — build a personal paper database with links
- ✓ Readwise Reader — highlights sync to a knowledge base
- ✓ Physical margin notes if reading on paper
Questions to ask every paper
- ✓ What problem does this solve?
- ✓ What is the main evidence it works?
- ✓ What assumptions does it rely on?
- ✓ What does it not address?
The 1-sentence test
After each section, close the PDF and write one sentence summarising it. If you cannot, reread. This forces active recall.
The figure-first trick
Look at every figure and caption before reading any prose. Figures carry 60% of the information in most ML papers — understanding them sets up the prose perfectly.
Spaced repetition
Log papers in Anki or a review system. Re-encounter a paper 1 day, 1 week, and 1 month later. Retention jumps from ~20% to over 80%.