AI + Safety · AI + Healthcare · Computational neuroscience
Research: How minds reason, fail, and learn
I study how humans and machines reason, and how biology can teach AI to reason better. The work sits where computational neuroscience, foundation models, and healthcare meet: not benchmarking for its own sake, but building systems that solve real problems where accuracy, safety, and trust matter.
Lab
Divergence Lab
Interactive work on reasoning, refusal, and faithfulness in language models: the gap between what a model reports and what it does.
Thread 01
Foundation-model safety and adversarial evaluation
How do frontier models fail when you push on them? Replicating the TEMPEST multi-turn attack against ten frontier models from eight vendors, with more than 97,000 API queries, found attack success of 96 to 100 percent on six of them, and showed that extended reasoning cut success from 97 to 42 percent. Companion work tests guardrail robustness, instruction adherence across 256 models, and a 1,554-prompt consensus-labeled bank for measuring how coding models respond to malicious-code requests.
Multi-turn adversarial attacks at frontier scale
Guardrail robustness and instruction adherence
Code safety: separating executable weapons from security knowledge
Thread 02
Chain-of-thought faithfulness
Does a model's stated reasoning match how it actually reached its answer? This thread measures divergence between thinking tokens and final answers in open-weight reasoning models, and shows that the answer depends heavily on how faithfulness is measured, including the sensitivity of the classifier used to judge it.
Thinking tokens versus answers in open-weight reasoning models
Classifier sensitivity in faithfulness evaluation
Abliteration methods compared across architectures
Chance favors the prepared mind.
Thread 03
Clinical AI, fairness, and safe deployment
AI used in healthcare has to be accurate, fair, and private at the same time. Work here includes a fairness audit of gender bias in language-model emergency-department triage, multi-model benchmarks for clinical-trial data extraction, specialized clinical-cardiology embeddings, and a systematic evaluation showing that vision-token masking alone cannot prevent protected-health-information leakage in medical document OCR.
Fairness auditing in emergency-department triage
Clinical-trial data extraction across model vendors
PHI privacy in medical document processing
Thread 04
Computational neuroscience and bioinformatics
The earlier work asks the same question of biological minds: machine learning and transcranial current stimulation in Parkinson's disease, genomic and brain-imaging features for Alzheimer's diagnosis, and the Minimotif Miner series in Nucleic Acids Research on short functional peptide motifs and their role in neurodegenerative disease.
Neuromodulation and motor learning in Parkinson's disease
Multimodal AI for Alzheimer's diagnosis
Minimotifs: a million peptide motifs and counting
Selected publications
Read the work
Twelve of thirty-plus. The full list, with abstracts and conference presentations, is on the CV.
- 01
- 02
- 03
- 04
- 05
- 06
- 07
- 08From API to Action: A Multi-Model Comparison of OpenAI, Anthropic, Google, and Meta LLMs for Clinical Trial Data Extraction
Bioengineering 13(7), 773 · 2026
- 09
- 10Benchmarking Multiple Large Language Models for Automated Clinical Trial Data Extraction in Aging Research
Algorithms 18(5), 296 · 2025
- 11Development of an AI Approach That Employs Genomic and Brain-Imaging Features to Improve the Diagnosis of Alzheimer's Disease
Alzheimer's & Dementia 18 · 2022
- 12Minimotif Miner 4: A Million Peptide Minimotifs and Counting
Nucleic Acids Research 46(D1) · 2018
Books and open source
Beyond the papers
Author of The Neuroscience of Artificial Intelligence: A Journey into the Minds of Machines and Humans and Healthcare Analytics and AI: Building Systems That Actually Work, and technical editor of Data Engineering with Alteryx. More than 80 open-weight model releases and datasets are on Hugging Face, with code on GitHub.
Two books on AI, neuroscience, and healthcare systems
80+ open-weight models, 55,000+ downloads
Two patents on human-in-the-loop clinical prediction
Work with me
Collaboration runs through DeepNeuro.
Research collaborations, advisory roles, applied AI projects, and speaking engagements run through DeepNeuro, my independent research practice. Students and UNLV colleagues can reach me directly at ryoung@unlv.edu.