Paper writing tips / nitpicks view markdown

A (growing) list of nitpicks that may help improve a paper (mostly focused on ML conference papers). This list is all opinions, but for brevity I’ll state them as facts.

In general, I highly recommend the book storytelling with data (it looks long, but it’s a very quick read!).


  • For every figure, think carefully about the message you intend it to convey and how it could be misinterpreted
  • Choose a single message to send with a figure and relentlessly remove all distractors
    • If a figure makes multiple points, split it into multiple figures
    • Clear figures ≠ beautiful figures
  • Use colors for curves/bars very sparingly. For ML plots with many curves, there are usually only a couple things to highlight (highlight only a few curves with color and leave the rest gray, e.g. this plot)
    • If curves have a natural ordering, then color them with a colormap, preferably a perceptually uniform one like viridis
    • Use diverging colormaps that are centered properly for values that are diverging like correlations (e.g. this plot)
    • Make figures colorblind-friendly (importantly, don’t use red & green together)
    • Keep colors consistent across different figures
  • Remove everything unnecessary
    • Remove top / right splines
    • Remove gridlines unless the precise values on a graph are useful for a reader
  • Label subpanels / individual points so they can be referenced directly in the text
  • Fonts should be readable: generally very similar size to the baseline fontsize of the paper
  • Figs should be in a vector format (e.g. pdf) so that their resolution stays sharp even when zoomed in. The only exception to this is when you’re plotting a ton of data, vector files can become very large and you should instead use a rasterized format with high resolution (e.g. png at 300 dpi).


  • Right-align numbers and keep rounding consistent (this makes it so that the decimal point lines up regardless of how large the numbers are)
  • Use bolding for emphasis
  • Show averages over different elements when possible


  • Here’s are some nice, brief notes on academic writing
  • Write in present tense
  • First-person is fine when it simplifies writing
  • Make your abstract really good: most people will only ever read the abstract
    • I prefer longer abstracts that give a description of the results
  • Introduction
    • AI papers often start with a grand opening
      • This is okay, but make the scope and limitations of your paper clear in the intro
    • Usually, the intro should have a compelling, easy-to-digest figure that motivates your approach; focus here on motivation rather than method details
  • Methods
    • Remove all jargon and math that is not strictly needed for your method
    • Give your method a name that is easy to remember and reference (ChatGPT is pretty good at coming up with these names; this website may also help)
  • Results
    • Start the results with a single figure/table with your key finding
    • Show ablation results with respect to each important hyperparameter/setting separately
    • Figure/table captions should only describe the data; main text should contextualize the results for the paper
  • Discussion
    • Don’t recap results
    • Address limitations and directions for future work
  • General
    • Text should move from simple to complicated, start with high-level motivation, then high-level solution before getting into detailed solution
    • Start with a paper outline and iteratively refine it to fill in details
      • This makes it easier to elicit feedback
      • Save the final polishing of writing / cleaning figures until everything is set
    • Writing should be clear and easy to skim