How to check if your research idea has already been done
Finding out six months in that someone published your idea last year is one of the most expensive surprises in research. A focused prior-art search before you start takes a day or two and usually settles the question. This guide walks through a method that works in any field.
Write the idea down as a testable claim
Vague ideas are impossible to search. Write one sentence with three parts: the problem, the mechanism you propose, and the outcome you expect. For example: "Allocating the KV-cache budget per request (mechanism) reduces peak inference memory (outcome) for long-context LLM serving (problem) without lowering accuracy."
Each part becomes a search axis. Most "has this been done?" answers turn out to be "the problem has been studied, the mechanism has been used elsewhere, but not this combination", and you can only see that if you search the parts separately.
Search the right places
Use the vocabulary of the field you are publishing in, then repeat the search with the vocabulary of neighbouring fields: the same mechanism often has different names in different communities.
- arXiv for recent preprints in computer science, physics, mathematics and quantitative fields; much new work appears there months before publication.
- OpenAlex and Semantic Scholar for broad coverage across fields, with citation links you can follow.
- Crossref for the published record and DOIs.
- Your field’s own index (for example PubMed for biomedicine) and the last two or three years of proceedings from the main venues.
Follow citations from the closest papers
Keyword search finds the obvious papers. The ones that scoop you are often found by following citations: take the three closest papers, look at what they cite (backward) and what cites them (forward). Repeat once with any new close paper you find. This is where most overlooked prior work turns up.
Compare overlap honestly
For each of the closest papers, compare it with your idea on three axes and write down the answer:
- Problem: do they address the same problem in the same setting?
- Method: do they use the same mechanism, or a different route to the same goal?
- Evaluation: do they measure the outcome you care about, on comparable data?
Argue against your own idea
Before deciding, write the strongest case that the idea is not new or will not work: the closest paper, the most likely confounder, and the result that would prove you wrong. If you can answer each point with evidence, the idea is probably worth pursuing. If you cannot, you have found what to change.
Decide: go ahead, adjust, or drop it
- Go ahead when the closest work differs on at least one axis that matters and you can say why the difference should change the outcome.
- Adjust when the overlap is large but a specific variation (a different setting, constraint or evaluation) is untested.
- Drop or reframe when a published paper already tests the same mechanism on the same problem and outcome.
Questions
How long should a prior-art search take?
For a single, well-defined idea, one to two days of focused searching usually finds the closest work. Broad ideas take longer; narrowing the claim first saves time.
Can a tool prove my idea is novel?
No. A search can only show the closest work it found. It reduces the risk of being scooped and tells you where the idea differs; the judgement is still yours.
What if I find a paper that does almost the same thing?
Compare it on problem, method and evaluation. A difference on one axis that changes the expected result can still be a contribution; frame your work around that difference.