Capstone: Write Your Own Paper
Pick a feasible first project in the field you choose, then work through a stage-by-stage paper workbook that saves as you go, exports a Markdown or LaTeX draft, and ends with a reviewer-style self-assessment.
What this capstone is for
Everything in Parts 1–3 comes together here: you will produce a real first paper draft in your chosen field. A completed, honest, well-evaluated study (posted on arXiv or submitted to a workshop) with released code is the strongest single piece of evidence in a PhD application, stronger than grades alone.
Be realistic: a first project usually takes 3–6 months part-time, and most first papers are rejected at least once. The workbook keeps you moving: each stage has a concrete output, and none of them requires a big GPU budget.
Feasible first projects
Good first projects answer a narrow question well with modest compute. Common, respected types:
| Type | Example | Why it works |
|---|---|---|
| Careful reproduction | reproduce a recent paper's main table; test whether results hold with more seeds or a new dataset | teaches the whole pipeline; reproducibility tracks and challenges publish these |
| New setting or population | evaluate known methods in a low-resource language, a new medical domain, or a different country's data | a clear gap; your local knowledge is an advantage |
| Analysis or evaluation study | when and why does method X fail? robustness, fairness or calibration across groups | no new method needed; rigorous analysis is valued |
| Benchmark or dataset | a carefully documented dataset for an under-served task (with ethics approval and consent) | high impact if done well; plan for data governance |
| Small method improvement | a simple change with a clear rationale, tested fairly with ablations | harder to show convincingly; needs strong baselines |
Try it: project ideas for your field
Choose your field and resources to see project patterns that suit a first paper. Then make one specific to a dataset or setting you know.
Interactive: this part runs in your browser.
The paper workbook
Work through the stages in order; your text is saved in this browser as you type (export regularly as a backup). Each stage shows the questions a reviewer will ask. Export a Markdown draft or a LaTeX skeleton for Overleaf at any time.
Interactive: this part runs in your browser.
A realistic timeline
| Weeks (part-time) | Stage | Output |
|---|---|---|
| 1–3 | question, reading, gap | one-sentence question; literature matrix of 15–30 papers |
| 4–5 | plan | written experiment plan (metrics, baselines, seeds, splits) reviewed by someone |
| 6–12 | experiments | baselines reproduced first, then your runs; results files and scripts in Git |
| 13–15 | analysis | tables and figures from code; ablations; error analysis |
| 16–19 | writing | full draft; feedback from two readers; revision |
| 20–24 | submission | arXiv preprint, code release, workshop or conference submission |
If you have no supervisor yet, ask a lecturer, a researcher whose paper you build on, or a community such as ML Collective for feedback on your plan and draft. Short, specific requests ("could you look at my two-page plan?") get replies.
Quiz
1. Which first project is most feasible with a laptop and a free GPU?
2. Why reproduce baselines before running your method?
3. The strongest single piece of evidence in a PhD application is usually…
📚 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.
- ML Collective ↗ open research community, mentoring and paper feedback
- S. Peyton Jones, How to Write a Great Research Paper ↗ talk and slides
- NeurIPS Paper Checklist ↗ what reviewers expect you to report