Reading Papers

---
name: reading-papers
description: Read research papers efficiently in passes, separating claims from evidence and assessing reproducibility. Use when extracting value from academic papers without reading every one cover to cover.
---

# Reading papers

Research papers are dense and mostly not worth reading in full; the
skill is triaging fast, extracting the claim and its support
efficiently, and judging whether to believe it. A pass-based approach
gets you the value of most papers in minutes and reserves deep reading
for the few that earn it.

## Method

1. **First pass: decide relevance in five minutes.** Title,
   abstract, figures, and conclusion: enough to know what
   the paper claims, roughly how, and whether it is relevant
   to your question (see literature-review's scope). Most
   papers you triage stop here (not relevant, or the
   abstract's claim is all you needed); this pass is
   filtering, not understanding.
2. **Second pass: get the contribution and method.** Read
   the introduction, the method at a conceptual level, and
   the results, skimming proofs and details: enough to
   understand what they did, what they found, and whether
   the approach is sound. This pass gives you the paper's
   actual content for the majority of relevant papers, in
   under an hour.
3. **Third pass: reproduce the reasoning, only when it
   matters.** For the few papers central to your work: read
   in depth, follow the derivations, question every step,
   and mentally (or actually) reconstruct the method: the
   level at which you could re-implement or critique it.
   Reserve this expensive pass for the load-bearing papers,
   not every citation.
4. **Separate claims from evidence.** What does the paper
   *claim*, and what does it actually *show*? The gap is
   where over-claiming lives: a result on one dataset
   claimed as general, a correlation described as
   causation (see experiment-design-basics), an improvement
   that is within noise. Read the results and their
   limitations, not the abstract's summary of them (see
   source-evaluation's claim-vs-evidence).
5. **Assess reproducibility and rigor.** Is there code and
   data? Is the sample size adequate and the comparison
   fair (baselines, ablations)? Are the limitations honestly
   stated? Has it been replicated or only published (see
   source-evaluation)? A striking result with no code, a
   tiny sample, and no replication is a hypothesis, not a
   finding: weight accordingly.
6. **Extract into your own notes and synthesis.** Capture
   the contribution, method, key result, and your
   assessment in your own words (see research-synthesis):
   the act of restating tests your understanding, and the
   notes feed the literature review without re-reading.
   Note how it connects to and disagrees with other work.

## Boundaries

- Pass-based reading triages efficiently but risks missing
  subtlety in a skimmed paper; the depth must match how
  much the paper matters to your decision or work. Do not
  deep-read everything, and do not act on a paper you only
  first-passed.
- A single paper is one data point; extraordinary claims
  need replication and convergent evidence before you
  build on them (see the replication ethic in
  source-evaluation). Published does not mean true.
- Peer review and venue are weak quality signals, not
  guarantees; strong papers appear in weak venues and
  flawed papers in strong ones. Judge the work, not just
  the stamp (see source-evaluation).