How to Read a Paper
paper reading, reading workflow, theorem reading, literature navigation
1 Why This Page Matters
Most readers lose time on papers because they go deep too early.
They start chasing a proof before they know:
- whether the paper is even worth a deep read
- whether the real blocker is notation, theorem structure, or missing prerequisites
- whether the main claim is theoretical, empirical, or mixed
This page gives the first-pass workflow for making that decision on purpose.
2 First-Pass Workflow
Use this order:
skim for structuremap the main claimsseparate theorem evidence from empirical evidenceidentify the real blockerdecide whether to read deeply now, later, or not at all
3 Core Workflow
3.1 1. Skim for structure
Before reading any proof carefully, identify:
- the problem the paper is addressing
- the section structure
- where the main theorem or main empirical result lives
- whether the heavy details are in the main body or appendices
The goal of this pass is orientation, not mastery.
3.2 2. Map the main claims
Ask what the paper is actually claiming.
Typical claim buckets:
theoretical: theorem, bound, guarantee, impossibility, characterizationempirical: benchmark gains, ablations, robustness behavioralgorithmic: method design, runtime, optimization behaviormixed: theorem plus experiments that illustrate or stress-test a regime
If you cannot name the paper’s core claim type, you are not ready for deep reading yet.
3.3 3. Identify the first blocker
At this point, decide what is blocking understanding most:
- dense symbols -> Notation Translation
- dense theorem statements -> Theorem Decoder
- too many lemmas, citations, or appendices -> Dependency Maps
- no clear next sequence of pages/papers -> Reading Trails
This branching step is what turns Paper Lab into a workflow instead of a loose set of pages.
3.4 4. Decide whether to read deeply now
A paper deserves deep reading now if:
- the main question matters to your current trail
- the core claim is readable after one of the branch workflows
- the missing prerequisites are small enough to repair
Otherwise, postpone it deliberately instead of half-reading it.
4 What To Mark Up
While reading, highlight:
- the main theorem or result
- the assumptions
- the notation you do not recognize
- the paper’s evidence for each major claim
5 Worked Example
Suppose you open a paper and see:
- one theorem in the introduction
- six appendices
- unfamiliar notation in the first displayed equation
- experiments that appear before the proof details
A good first-pass reaction is:
- skim the abstract, introduction, and theorem statement
- decide whether the main claim is theoretical or mixed
- notice that notation is already blocking the theorem
- open Notation Translation before decoding the theorem
- then open Theorem Decoder if the statement is still dense
- only after that decide whether appendix proofs are worth deep time now
That is much better than jumping straight into Appendix C.
6 Common Failure Modes
- reading linearly from page 1 to the appendix without deciding what matters
- spending thirty minutes chasing a symbol before deciding whether the paper deserves deep reading
- treating theorem statements and experiments as the same kind of evidence
- assuming the first hard-looking theorem is automatically the right place to start
- postponing every hard part until the reading session collapses into skimming
7 Sources and Further Reading
- How to Read a Paper -
First pass- still the best short staged-reading workflow. Checked2026-04-25. - How to Read a Research Paper -
Second pass- useful for moving from structure to critique and synthesis. Checked2026-04-25. - CS167 Technical Paper Reading Guide -
Second pass- especially useful when papers mix theory, experiments, and implementation details. Checked2026-04-25. - MIT Mathematics for Computer Science: Introduction and Proofs -
Second pass- useful when the blocker is proof structure rather than paper context. Checked2026-04-25. - NeurIPS Paper Checklist -
Paper bridge- current venue-side reminder to read assumptions, limitations, and evidence boundaries explicitly. Checked2026-04-25.