Becoming an AI Researcher
Topic 13 · Part 3 · Writing & publishing · ⏱ about 5 h

Figures, Tables, LaTeX & Overleaf

Figures that carry the argument, honest charts with uncertainty, clear tables, LaTeX and Overleaf, venue templates, BibTeX, and the pre-submission checklist.

Show:
In this topic
Beginner

Figures carry the argument

Figure 1 (often on page 1 or 2) should make the main idea or result understandable at a glance: an overview diagram, or the key result. Every figure needs a takeaway: say it in the caption ("Adapters degrade less than full fine-tuning as labels decrease").

  • Show uncertainty: error bars (state whether std, standard error or 95% CI) or shaded bands over seeds.
  • Label axes with units; use readable font sizes (as large as the caption text); avoid tiny legends; place labels directly on lines where possible.
  • Use colour-blind-safe palettes, and do not rely on colour alone (markers, line styles).
  • Start bar charts at zero; for line charts choose ranges that show differences honestly; never truncate axes to exaggerate a gain.
  • Export vector graphics (PDF) for plots; keep the code that makes every figure, run from the results files.
Try it

Try it: fix a misleading figure

Toggle the common problems on and off and see how the same results can look decisive or honest.

Interactive: this part runs in your browser.

Beginner

Tables

Use the booktabs style (no vertical lines, three horizontal rules). Put the main result first; group related columns; give units and the direction of better (↑/↓); bold the best result only when the difference is meaningful (or mark statistically indistinguishable results); report mean ± std or CIs; make the caption explain abbreviations. Generate tables from code, never by retyping numbers.

Beginner

LaTeX, Overleaf and BibTeX

Almost all AI venues require LaTeX with their own template. Overleaf (free, in the browser) lets co-authors edit together, with history and comments. Practical habits:

  • Download the official template for the venue and year; never change margins, font sizes or spacing (papers are desk-rejected for it). Anonymise for double-blind review.
  • One file per section (\input{sections/intro}); macros for repeated terms (\newcommand{\method}{FastTune}).
  • BibTeX: export entries from Zotero (Better BibTeX keeps keys stable); cite published versions rather than arXiv where they exist; check every entry (authors, year, venue).
  • Use \cref or \autoref for references to figures and sections; siunitx for numbers; booktabs for tables.
Beginner

Before you submit

  • Page limit respected (main text), references and appendix where allowed; anonymised (no names, acknowledgements, identifying links; anonymous code links via services such as anonymous GitHub).
  • Every claim in the abstract and introduction is supported in the results.
  • All figures readable in greyscale and when printed; captions self-contained.
  • Numbers in text match the tables (generate both from the same files).
  • Checklists (reproducibility, ethics) completed honestly; supplementary material uploaded.
  • A colleague has read the whole paper; spell-check done; all co-authors approved the final version.
Code

Code: a publication-quality figure

Mean and 95% CI over seeds, colour-blind-safe, vector output.

Python · matplotlib
import numpy as np, matplotlib.pyplot as plt
plt.rcParams.update({"font.size": 9, "axes.spines.top": False, "axes.spines.right": False})
labels = np.array([50, 100, 200, 500, 1000, 2000])
runs = {"Full fine-tuning": full_runs, "Adapters (ours)": adapter_runs}      # each: (seeds, len(labels)) array
fig, ax = plt.subplots(figsize=(3.3, 2.3))                                 # one column wide
for (name, R), c, m in zip(runs.items(), ["#0072B2", "#D55E00"], ["o", "s"]):
    mu = R.mean(0); ci = 2.776 * R.std(0, ddof=1) / np.sqrt(len(R))         # t(0.975, 4) for 5 seeds
    ax.plot(labels, mu, marker=m, color=c, label=name); ax.fill_between(labels, mu - ci, mu + ci, color=c, alpha=0.2)
ax.set_xscale("log"); ax.set_xlabel("labelled examples"); ax.set_ylabel("macro-F1 (%) ↑")
ax.legend(frameon=False); fig.tight_layout(); fig.savefig("fig_labels.pdf")   # vector output for LaTeX
Open in playground ↗
Before it runs, the course makes example results: 5 seeds × 6 label budgets for each method.
Check yourself

Quiz

1. What should every figure caption contain?

2. A bar chart of accuracies starts its y-axis at 80%. The risk is…

3. Changing the template margins to fit more text…

4. Best source of numbers for tables in the paper?

5. For double-blind review, your code link should be…

Go deeper

📚 Further reading

Trusted, mostly free sources: official documentation, well-known textbooks and courses. When you want to confirm or go deeper on anything here, start with these.

Ready