01
Behavioral: locate the gap
Scaled, replicable measurement of where safety training holds and where it doesn’t. Identical requests, different surface forms — natural language refused, executable code complied with.
DIVERGENCE LAB / Lee Business School, UNLV · coming soon
Reasoning · Evaluation · Measurement
A model’s account of its reasoning and its actual reasoning are two different traces.
We measure the gap. Divergence Lab studies reasoning in artificial systems — what it is, when it’s real, and when it’s confabulated — using the experimental methods developed for biological minds.
Research program
Failure is the instrument, not the object.
Neuroscience learns about vision from illusions, memory from amnesia, language from aphasia. We learn about machine reasoning from the places it comes apart — because breakdown is where mechanism becomes visible. The three lines below are sequential: each answers a question the one before it raises.
02
Trace: read what the model says it did
Whether thinking-token traces reflect the computation that produced the answer, and how much of measured faithfulness is an artifact of the classifier doing the measuring.
03
Mechanism: test whether it’s one norm or two
If code safety is a separable direction in activation space rather than a general principle applied to new inputs, the model has two unlinked mechanisms — not one rule it failed to generalize.
Open artifacts
Everything we measure with, we release.
Benchmarks, prompt banks, and model weights ship publicly. An evaluation you can’t reproduce isn’t a measurement — it’s an anecdote.
55K+
Model downloads
80+
Open-weight releases
97K+
Adversarial queries logged
256
Models evaluated
Weights and datasets on Hugging Face and Ollama · code on GitHub · publications on the research page
Working with the lab
You can run a real experiment in fifteen weeks.
This work parallelizes unusually well. A single model family, a single stimulus set, a single replication — each is a self-contained study that composes into a paper. If you can write Python and read a results table honestly, you can contribute.
- Open to MS Data Analytics, MIS, and capstone students at UNLV
- Authorship on what you produce — not a line in the acknowledgments
- No IRB delay: model evaluation isn’t human-subjects research, so you start in week one
- You’ll leave with a public artifact — a dataset, a benchmark, or a preprint with your name on it
People
Richard J. Young, Ph.D.
Principal investigator. Assistant Professor-in-Residence, Information Systems, Lee Business School, UNLV. Ph.D. in computational neuroscience; thirty-plus publications across AI safety evaluation, clinical AI, and neuromodulation. The lab’s methods come from the first of those; its urgency comes from the second.
Collaborations with G. D. Moody (UNLV, code safety) · A. M. Matthews (clinical embeddings) · B. Poston (neuromodulation)
Richard J. Young, Ph.D.
Director
AI safety, clinical AI, and foundation model evaluation. Assistant Professor-in-Residence, Information Systems, UNLV; Senior AI Research Scientist, UnitedHealth Group.
Alice Matthews, Ph.D., RDCS
Research Scientist
Cardiac sonographer and neuroscientist bridging cardiac imaging with clinical AI. Co-author on CardioEmbed.
Lucky
Lab Mascot
Morale, quality assurance, snack disposal.
Founding Graduate Student
Graduate Research Assistant · open seat
First graduate seat in the lab — foundation model evaluation or clinical AI. Your name replaces this card.
Founding Undergraduate RA
Undergraduate Research Assistant · open seat
Hands-on research for a UNLV undergrad who finishes what they start. No prior experience required.
Clinical Research Student
Research Assistant · open seat
For a student from the health sciences side — ultrasound, nursing, medicine — curious about AI in the clinic.
Divergence Lab · Lee Business School · University of Nevada, Las Vegasdeepneuro.ai · CV · ORCID