When Graph Theory Meets AI: The Strangest Skill Assessment Questions We've Seen
Our AI generates skill assessment questions for whatever professionals add to their profiles. Most of the time it's Docker, Agile, and SQL. Sometimes it's Convexified Modularity Maximization -- at beginner difficulty.
Obvelum lets professionals list any skill on their profile -- from "React" to "stakeholder management" to things we had genuinely never heard of. When a skill gets verified, our AI generates multiple-choice questions calibrated to a difficulty level. Most of the time, this works exactly as expected.
Then one day, our admin console lit up with a cluster of assessment questions that stopped us cold. A single user had built a profile that read less like a tech resume and more like a PhD qualifying exam in network science.
The Questions
Here are six real assessment questions generated from that user's profile -- each marked as difficulty level 1 (beginner). We have not edited them.
Privacy note: profiles on Obvelum are anonymous by design. We can see aggregate skill data and AI-generated questions, but we have absolutely no idea who this person is. We cannot. That is the whole point.
What is the primary motivation for using a convexified version of the modularity objective in community detection?
In a degree-corrected stochastic block model (DC-SBM), what does the degree correction primarily account for?
In the Girvan-Newman algorithm, which edges are iteratively removed to detect community structure?
In the Affiliation Graph Model (AGM), how are edges between nodes generated?
In the Label Propagation Algorithm for community detection, how does a node determine its community label?
Which of the following is a known limitation of optimizing modularity (Q) for community detection?
That Is Actually the Point
The AI had clearly been trained on academic literature. It recognized "Convexified Modularity Maximization" as a real, citable concept and generated a question with three plausible distractors and a defensible correct answer -- all at what it estimated was introductory difficulty for someone learning the topic.
That's the system working correctly. Obvelum doesn't curate a list of "valid" skills. If you know it, you can list it. The AI adapts to whatever you bring, not the other way around.
The user with the graph theory background sailed through every question. Their profile went live. Somewhere, a hiring manager is reading about stochastic block models for the first time.
What We Learned
- The long tail of human expertise is genuinely long. Our AI handles it.
- Difficulty "1" is relative to the domain, not to the general population.
- Academic graph theorists exist, they job hunt, and they appreciate not being asked to describe a linked list instead.
- Running a prod query and getting back "Convexified Modularity Maximization" at 11pm is an experience.
If you have a niche skill that no other platform has ever asked you about -- add it. We'll generate an honest assessment of it. No guarantees about what the AI does next.