From Idea to Research Question
Where good ideas come from, turning an interest into a sharp, answerable question and hypothesis, the FINER test, scoping a project to months, and the one-sentence contribution.
Where ideas come from
- Limitations sections and "future work" of recent papers: authors tell you what they could not do.
- Failures you observe: a method that breaks on your data, a metric that disagrees with human judgement.
- Combining fields: a technique from one area applied where it has not been tried, with a reason to think it fits.
- Your edge: a language, a domain, a dataset or a community you know well.
- Questions behind results: "it works, but why?", "does it still work when…?", "what is it really learning?"
Keep an ideas log: one line per idea, with the date and where it came from. Most will be bad; the log lets the good ones surface and combine over time.
From topic to question to hypothesis
A topic is an area ("LLMs for healthcare"). A research question is something your experiments can answer ("Does retrieval of clinic policies reduce unsupported answers by a small open model compared with prompting alone?"). A hypothesis is your predicted answer with a reason ("Yes, by at least half, because most errors are missing facts rather than reasoning errors"), stated so that a result could show it false.
A useful template: In [setting], does [intervention] change [measurable outcome] compared with [baseline], and why? Each bracket forces a decision you must make anyway: the data, the method, the metric, the comparison.
The FINER test
| Criterion | Ask yourself |
|---|---|
| Feasible | Do I have the data, compute, skills and time (weeks to months)? What is the smallest version I could finish? |
| Interesting | Would the answer surprise or matter to people in the field (and to me, for months)? |
| Novel | Has it been answered already? What is different about mine? |
| Ethical | Data consent and licences, privacy, potential harms, approvals needed? |
| Relevant | Does the answer change what someone does or believes? |
Originally from clinical research, FINER is a quick filter for any AI project. Scope ruthlessly: a narrow question answered convincingly is worth far more than a broad one answered weakly.
Try it: sharpen a research question
Start from a vague idea and make it answerable by choosing the setting, intervention, outcome and baseline. The FINER checks update as you go.
Interactive: this part runs in your browser.
The one-sentence contribution
Before you run the main experiments, write the sentence you hope will open your abstract's results: "We show that X, which matters because Y." Then list the figure or table that would prove it and the experiment that would produce it. If the sentence is boring even when everything works, change the question now, not after three months.
Plan for the likely outcome too: if the hypothesis is false, is the negative result still informative? Questions that teach you something whichever way the result goes are the safest for a first project.
Quiz
1. Which is a research question rather than a topic?
2. A hypothesis should be…
3. The F in FINER stands for…
4. Why prefer a question where either outcome is informative?
5. Where can you reliably find open problems the authors themselves admit?
📚 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.
- A. Karpathy, "A Survival Guide to a PhD" ↗ choosing problems, writing, and PhD life
- J. Eisner, Advice for Research Students ↗ a curated collection of advice on reading, writing and research