What kind of CS research project is worth doing before applying to grad school?

Pick a small question you can answer carefully, rather than trying to make a project sound groundbreaking. For example, test whether outdated documentation makes an LLM coding assistant less reliable, or whether errors build up when it works across several sessions. Define the task, compare a few conditions, and report what you measured, including limitations.

A clear method and honest evaluation will usually make a stronger portfolio piece than a complicated stack with no convincing results.

A practical route is to build something around a real use case, then evaluate it instead of stopping at a demo. A retrieval system for a particular set of documents could work, but compare it with a simpler baseline and measure things like answer accuracy, unsupported claims, and how often it retrieves relevant material.

The topic itself does not have to be flashy. A reproducible experiment and a clear explanation of what did and did not work are the useful parts.

Before settling on a topic, read a few recent papers in an area you actually care about. For each one, note what question it asks, how it tests the idea, and what limitations or unanswered questions the authors mention. A modest extension of an existing study can be a sensible starting point.

If you can, ask a faculty member or research group for feedback early. They can help you narrow the question and avoid spending months on a project that is too broad.

Don’t treat publication as the only sign that a project is good. A well-documented replication, benchmark, or careful negative result can show research skills even if it does not become a paper. Keep the code, data choices, and evaluation steps understandable so someone else can check your work.