[reference] — from The Craft of Research and standard lit-review methodology. The hardest, most underrated part of research: choosing what to work on. A well-chosen question is half the paper; a badly chosen one wastes months no matter how well you execute. This note is how to find a good problem and how the literature review proves it’s worth doing.

The kid version first

Everyone thinks the hard part of research is answering the question. It’s not — it’s finding a good question in the first place. A good research question is like a spot to dig for treasure that is all three of these at once:

  • Nobody has dug there yet (it’s novel — not already answered),
  • There’s probably treasure (it matters — someone will care about the answer), and
  • Your shovel can actually reach it (it’s feasible — you can tackle it with your time, skills, and equipment).

Digging where someone already dug = you rediscover their treasure (not novel). Digging where there’s no treasure = wasted effort (not significant). Digging toward the Earth’s core with a plastic spade = impossible (not feasible). Most bad projects fail one of these three before any work is done. And the map that tells you where people have already dug is the literature review.

Good questions are the scarce resource

You will have no shortage of energy to do research and a chronic shortage of good questions to point it at. Experienced researchers are valuable largely because they have a taste for good problems. That taste is learnable — it comes from reading a lot (reading papers) and knowing where questions hide.

Where good research questions come from

Questions rarely arrive as lightning bolts; you mine them:

  • The “Limitations” and “Future Work” sections — authors literally list what their paper doesn’t solve. This is the single richest vein: they’ve handed you open problems.
  • “Does it still hold if…?” — take a result and change one thing: a different dataset, a harsher condition, a bigger scale, a real-world setting instead of a lab one. Replication under new conditions is real, publishable research (and often reveals the original was overclaimed).
  • Cross-pollination — apply a method from field A to a problem in field B (“this technique from NLP might help this robotics control problem”). Much innovation is transfer, not invention.
  • Where papers disagree — two papers claim opposite things; why? Resolving a contradiction is a great contribution.
  • Where a method breaks — you notice (in a Pass-3 deep read) a hidden assumption; what happens when it’s violated?
  • Your own itch — a real problem you hit while building something. Systems/engineering research often starts here: “I needed X, nothing did it well, so I built and evaluated it.”

For a systems-engineering student researching across AI/ML, robotics, and chips, the richest sources will usually be “does this ML method work on my real physical/hardware system?” and “I built a system to do X, here’s the design and evaluation.”

The literature review — what it’s really for

Beginners think a literature review is “a summary of everything written on my topic.” It’s not. A literature review is an argument that (a) establishes what’s already known, and (b) reveals the specific gap your work fills. It exists to prove your question is novel (nobody dug here) and to situate your contribution (which existing bricks you build on).

How to do one well:

  1. Search systematically — start from a survey + seminal papers, follow the citation map backward and forward. Search multiple term variations (fields use different words for the same idea).
  2. Organize by theme, not chronology — group papers by approach or sub-problem (“methods that do X vs methods that do Y”), not “Smith 2019 did… then Jones 2020 did…”. A chronological list is a red flag; a thematic synthesis shows understanding.
  3. Synthesize, don’t enumerate — say what the body of work collectively shows and where it agrees/disagrees, citing groups of papers together. You’re drawing a map, not writing book reports.
  4. Name the gap explicitly — end with the sentence the whole review was building toward: “However, no prior work has addressed ___ — which is what this paper does.” That gap is your contribution’s justification.

A gap is a known unknown — something the field would benefit from knowing but doesn’t yet. Not every gap is worth filling (some are boring); the good ones are gaps whose answer matters.

Scoping the question — the Goldilocks zone

Once you have a candidate question, right-size it. The classic checklist is FINER:

  • Feasible — can you actually do it with your time, skills, data, and equipment? (the one students most overestimate)
  • Interesting — will anyone care about the answer?
  • Novel — not already answered (this is what the lit review checks).
  • Ethical — can it be done responsibly? (ethics)
  • Relevant — does it advance the field / matter to someone?

And apply two blunt tests:

  • The Goldilocks test — too broad (“solve robot navigation”) is a career, not a paper; too narrow (“tune this one hyperparameter”) is trivial. Aim for a question you can answer convincingly in one focused piece of work.
  • The “so what?” test — imagine you’ve finished and gotten a clean answer. Say the answer out loud. Does anyone care? If “so what?” has no good reply, re-scope before you start.

Converting a fuzzy topic (“I’m interested in drones”) into a sharp question (“does controller X reduce landing error vs the standard PID under wind gusts >5 m/s?”) is most of the battle — the sharp version is testable and finishable.

Staying current (an ongoing habit, not a one-time task)

The frontier moves; keep up cheaply: arXiv subject alerts, Google Scholar alerts (for a topic or when a key paper gets cited), follow the key researchers/labs (their new work + who they cite), and skim the accepted-papers lists of the field’s top venues each cycle. Fifteen minutes a week keeps your map current and often surfaces your next question.

Common traps (all avoidable before you start)

  • Reinventing the wheel — you skipped/rushed the lit review and built something that already exists. The most demoralizing and most preventable failure.
  • Too broad — you’ll never finish; nothing gets a convincing answer.
  • Not actually novel — “new to me” ≠ “new to the field.” Only the literature can tell you.
  • Not feasible — the experiment needs data/compute/equipment/time you don’t have. Check feasibility before committing.
  • No “so what?” — technically fine, but nobody cares.

Key insight

The scarce resource in research is a good question, and finding one is half the work. A good problem is simultaneously novel (nobody’s answered it), significant (the answer matters), and feasible (you can actually do it) — questions hide in papers’ Limitations/Future-Work sections, in replication-under-new-conditions, in cross-field transfer, and in your own build-it pain. The literature review is not a summary but an argument that reveals the gap you’ll fill — organized by theme, synthesized (not enumerated), ending in the explicit “no prior work has done ___” sentence. Scope with FINER and the “so what?” test, and turn a fuzzy topic into one sharp, testable question before you invest a single day of real work.

↑ Back to Research & Paper Writing