DEV Community

Jason
Jason

Posted on

A four-branch map for reading research papers

Graduate students and researchers already know how to mark up a PDF. The harder part is opening that paper three months later and still knowing what mattered: the problem it attacked, how the method is supposed to work, what evidence would change your mind, and where the authors say the work stops.

Use one map per paper, with the same four branches every time: Problem, Method, Evidence, and Limitations. Fill the skeleton by hand in a text file, then turn the outline into a foldable map if you want one, and check it against the PDF.

Four questions, one branch each

Read with these questions in the margin. If a paragraph does not help answer one of them, it probably belongs as a detail under a branch, or it does not belong in the note.

Problem. What situation is broken or incomplete, and for whom? Quote or tightly paraphrase the authors' own statement of the gap, and name the prior approach they are arguing against. "We need a better model" is too thin. "Recurrent encoder-decoder models preclude parallelization within a training example, and a recurrent layer has a maximum path length of O(n)" is usable, because a later paper can be checked against that same gap.

Method. What is the actual proposal, in one or two sentences a colleague could understand? Then list the pieces you would have to reimplement or cite: the architecture, the loss, the data, the training setup, the evaluation protocol. Keep "the idea" and "the recipe" on separate sub-nodes, or you cannot compare two papers that share a slogan and differ in the recipe.

Evidence. What was measured, on what, and against what baseline? Write the metric and the dataset or task, and point to the table or figure you would look up if you had to defend the claim. Leave result numbers out of the note until that table is open. A node that says "improves translation quality, see Table 2" ages better than a number you half-remember.

Limitations. Start with what the authors concede: scope of the tasks, compute, assumptions, failure cases. Then add one line of your own: what this paper cannot be used to claim.

A paper will not always use these headings. Look in the introduction and related work for the problem, in the method section for the recipe, in experiments for evidence, and in the discussion, analysis, or a comparison table for limits. If a section is missing, leave the branch in place and write "not reported" plus where you looked.

A reusable outline

Keep this as Markdown under each paper's title. The mermaid tree is the same structure.

Paper: [title, authors, year]
- Problem
  - Gap (authors' terms)
  - Prior approach they argue against
  - Who is affected / in what setting
- Method
  - Core proposal (one or two sentences)
  - Pieces: architecture, loss, data, training, evaluation
  - Anything you would have to reimplement
- Evidence
  - Metric + dataset or task
  - Baseline
  - Pointer to table/figure (numbers only after you open it)
  - What result would change your mind
- Limitations
  - What the authors concede
  - Scope of the experiments
  - What you will not claim on the basis of this paper
Enter fullscreen mode Exit fullscreen mode
mindmap
  root((Paper))
    Problem
      Gap in the authors' terms
      Prior approach they argue against
      Setting
    Method
      Core proposal
      Recipe and setup
    Evidence
      Metric and task
      Pointer to table or figure
    Limitations
      Authors' caveats
      What you will not claim

One paper, one map. For a literature review, make a second map whose branches are questions (gap, method family, evidence type, open limit) and whose child nodes are paper titles. Do not merge two papers into one tree on the first pass.

Worked example: Attention Is All You Need

The filled branches below are a note-taking illustration, not a substitute for the paper, and not a report of a tool run. The source is Vaswani et al., 2017, "Attention Is All You Need" (arXiv:1706.03762). Claims follow the abstract, Sections 3, 4 and 6, and Tables 1, 2 and 4. Translation scores are from Table 2; if you cite a number, confirm it against the table in the copy you are reading.

Problem. The abstract states that the dominant sequence transduction models are recurrent or convolutional encoder-decoder networks, and that the best of them also connect the encoder and decoder through attention. Section 1 explains that a recurrent model computes each hidden state from the previous one, which precludes parallelization within a training example. Table 1 lists a maximum path length of O(n) for a recurrent layer and O(1) for self-attention; Section 4 argues that shorter paths make long-range dependencies easier to learn.

Method. The Transformer is an encoder-decoder based solely on attention, with no recurrence or convolution (abstract; Section 3.5). In the base model in Section 3.1, encoder and decoder are stacks of N = 6 layers. Each encoder layer has multi-head self-attention and a position-wise feed-forward network, plus residual connections and layer normalization. The decoder adds multi-head attention over the encoder output and masks future positions. Section 3.5 adds positional encodings to the input embeddings so the model can use token order.

Evidence. Section 5.1 and Table 2 use WMT 2014 English-to-German and English-to-French, with BLEU on newstest2014. Table 2 lists 28.4 BLEU for Transformer (big) on English-to-German and 41.8 on English-to-French. Section 6.3 and Table 4 report English constituency parsing on Section 23 of the WSJ Penn Treebank, including a 4-layer model and a semi-supervised setting. Point the evidence node at Tables 1, 2 and 4.

Limitations. Table 1 gives self-attention a per-layer cost of O(n^2 d), against O(n d^2) for a recurrent layer. Section 4 says this is still faster than recurrence when sequence length n is smaller than representation size d, which is often true for sentence representations in machine translation, and flags restricted attention as future work for long sequences. The measured tasks are the two WMT 2014 translation benchmarks and English constituency parsing. The conclusion plans to apply the model to other tasks and modalities; those results are not in this paper. A caveat to attach: do not treat the paper as having measured every setting in which the architecture is now used.

If a generated map invents a limitation the authors did not discuss, or turns a "we plan to try X" sentence into a result, fix the node or delete it and go back to the PDF.

From bullets to a map you can fold

Once the four branches exist as headings and bullets, you already have a usable note. A map is optional. It helps when you want to collapse Evidence while you reread Method, or to keep the same outline in a Markdown vault.

Paste the outline into a text to mind map box and generate from the text you supplied. Short prompts such as "make a map of this paper" can fall back on what a model already knows, which is a poor fit for a paper in hand. Then walk one Evidence claim and one Limitations sentence back to the PDF.

If you still have the file, a pdf to mind map pass is usually cleaner than copying text out of the PDF, which often breaks headings and wraps lines mid-sentence. After a file upload, check that Methods and Results did not get merged, and that a limitations or discussion section still has its own branch. The same tool pages say you can try a short text paste without an account, with a 5,000-character cap, and that saving, editing, and exporting a map requires signing in. That cap will not hold a whole paper, so extract the four branches first.

Treat any generated tree as a draft of your note. Rename nodes that drifted. If a branch is missing, add "not reported" in the source Markdown and generate the map again from that outline. Official tool pages say renaming is available on the free plan, while adding or deleting nodes on the map may require a paid plan, so a missing branch is safer to fix in the text you paste than inside the editor. Export Markdown if you want the same outline in Obsidian or another editor. The downloaded file is a copy and will not sync later edits by itself.

When the next paper arrives

Fill a new four-branch map before you update any review map, then compare Method recipes and Limitation lines side by side. Papers that share a slogan often disagree in the evaluation protocol; that disagreement belongs under Evidence.

Pick one paper already on your desk. Write the four branches in bullets, in one sitting, without generating anything. Then, if you turn the outline into a map, verify one evidence pointer and one limitation against the PDF before you file the note.

Top comments (0)