Writing and sharing your results
Research that nobody can read or check does not help anyone. Learn the structure of a paper, clear figures, honest limitations, and how to share through preprints, code, blogs and workshops.
In 60 seconds
Writing and sharing your results
Research that nobody can read or check does not help anyone. Learn the structure of a paper, clear figures, honest limitations, and how to share through preprints, code, blogs and…
The structure of a paper
| Section | What it answers | Tip |
|---|---|---|
| Title | What is this about? | Specific and plain. Say the finding if you can. |
| Abstract | Problem, method, main result, why it matters | Four or five sentences. Write it last, then rewrite it. |
| Introduction | Why should I care? What did you do? | End with a short list of your contributions. |
| Related work | How does this fit with others' work? | Be fair to others. Explain the difference, don't just list papers. |
| Method | What exactly did you do? | Enough detail that someone could rebuild it. |
| Experiments | How did you test it, and what happened? | Data, baselines, metric, seeds, then results and ablations. |
| Limitations | Where does it not work, or not apply? | Be specific and honest. |
| Conclusion | What should the reader remember? | Short. No new claims. |
| Appendix | All the details | Hyperparameters, extra results, compute used. |
Clear figures and tables
- One figure, one message. Write that message as the caption's first sentence.
- Label the axes with names and units. Make text readable when printed small.
- Show the spread: error bars or shaded regions across seeds, and say in the caption what they mean (for example, standard deviation over 5 seeds).
- Use the same axis range when readers should compare two plots.
- Don't rely only on red and green. Use colours that people with colour blindness can tell apart, and different line styles or markers.
- In tables, mark the best result, and include the number of seeds.
- "Figure 3: Results."
- "Figure 3: Back-translation improves accuracy on all three Indian languages we tested. Bars show mean over 5 seeds; error bars show one standard deviation."
Honest limitations
- "More experiments could be done."
- "Our method may have some limitations."
- "We tested only models up to a small size; larger models may behave differently."
- "Our test set has news text only; results on social media text may differ."
Ways to share
| Where | What it is | Good for |
|---|---|---|
| arXiv preprint | A free, public paper server. Papers appear quickly, without full peer review. | Getting a dated, citable public version of your paper. First-time submitters may need an endorsement; check arXiv's current rules. |
| Open code | A public repository with your code | Letting others reproduce and build on your work. |
| Blog post | A shorter, friendlier write-up | Reproductions, tutorials, negative results, lessons learned. |
| Workshop | A smaller, focused event, often held alongside a larger conference | Early or in-progress work, meeting people in your area, getting feedback. |
| Conference or journal | Peer-reviewed venues | Mature, complete work. Read each venue's rules on preprints and anonymity. |
Open code that others can actually run
- 1
README
What the project does, how to install, how to run the main experiment with one command. - 2
Pinned versions
A requirements file with the library versions you used. - 3
Configs and seeds
The exact settings and seeds behind every number in the paper. - 4
Data instructions
Where to get the data and how to prepare it. Respect the data's licence. - 5
Licence
Choose an open licence so others know what they may do with your code.
Watch and read more
Lab
A one-page results note: title, abstract, one table or figure with a strong caption, and specific limitations.
The problem
Use the results from your seed-variance lab, your reproduction, or any small experiment of your own.
- Write a specific title that states the finding.
- Write a four- or five-sentence abstract: problem, what you did, main result, why it matters.
- Make one table or figure. Show the mean and spread over seeds, and write a caption whose first sentence is the message.
- Write a short "Setup" paragraph: data, split, baselines, metric, number of seeds, how settings were chosen.
- Write at least two specific limitations.
- Write a README outline for the code: install, run, where the numbers come from.
You are done when
Hard questions
Try to answer before you reveal. If you can answer these, you understood the lesson.
Q1Your result is clearly positive on two datasets and unclear on a third. How do you write the abstract?Reveal
Say both. For example: "improves accuracy on two of three datasets; on the third the difference is within seed variation." Hiding the third result would overclaim, and readers who try it may lose trust in your work. The honest version is still a useful finding.
Q2A reviewer says your figure's error bars overlap heavily. What should you do?Reveal
Take it seriously. Check whether your claim depends on a gap smaller than the spread. Options: run more seeds, use a paired comparison over the same seeds, or soften the claim to "no clear difference". Changing the plot to hide the overlap is not an option.
Questions people ask
Should I post on the arXiv before or after peer review?
Many AI researchers post a preprint when they submit, or when the work is complete. Some venues have rules about preprints and anonymity, so read the venue's policy first. Discuss the timing with your co-authors or mentor.
Is a negative result worth writing up?
Yes, if the experiment was clean. "This popular idea did not help in this setting" saves others time. A blog post, a workshop paper or a short report are all good homes for it. Explain carefully what you tried so readers trust it.
My English is not perfect. Will that hurt?
Clear and simple is better than fancy. Short sentences, one idea per paragraph, and consistent terms go a long way. Ask a friend to read it. Writing tools can help with grammar, but check the result keeps your meaning, and follow the venue's rules on AI use.
Who should be an author?
People who made a real contribution to the ideas, experiments or writing, and who agree to be responsible for the paper. Discuss authorship early and openly with your collaborators, not at the last minute. Thank others who helped in the acknowledgements.
Lesson test
5 questions. Get 3 right (60%) to pass and complete this lesson.