Publications, studies, and talks at the edge of neuroscience and AI.

EQUITRIAGE: A Fairness Audit of Gender Bias in LLM-Based Emergency Department Triage

Young, Richard J.

A systematic fairness audit of large language model behavior when used for emergency department triage decisions, focused on gender bias in acuity scoring and treatment recommendations.

A Validated Prompt Bank for Malicious Code Generation: Separating Executable Weapons from Security Knowledge in 1,554 Consensus-Labeled Prompts

Young, Richard J.

A 1,554-prompt consensus-labeled benchmark distinguishing prompts that elicit executable malicious code from prompts that probe defensive security knowledge, enabling sharper evaluation of LLM misuse-resistance without conflating the two failure modes.

Why Models Know But Don't Say: Chain-of-Thought Faithfulness Divergence Between Thinking Tokens and Answers in Open-Weight Reasoning Models

Young, Richard J.

Open-weight reasoning models often surface relevant information inside their thinking tokens that does not propagate to the final answer. This work measures that divergence at scale and characterizes when chain-of-thought traces stop being a faithful explanation of the model's output.

Lie to Me: How Faithful Is Chain-of-Thought Reasoning in Reasoning Models?

Young, Richard J.

An empirical evaluation of chain-of-thought faithfulness across modern reasoning models, testing whether the stated rationale actually reflects the computation that produced the answer, and quantifying systematic failure modes.

Measuring Faithfulness Depends on How You Measure: Classifier Sensitivity in LLM Chain-of-Thought Evaluation

Young, Richard J.

Faithfulness scores for chain-of-thought reasoning depend heavily on the classifier used to judge them. This work shows how classifier sensitivity drives apparent disagreements across faithfulness studies and proposes more robust evaluation protocols.

Comparative Analysis of LLM Abliteration Methods: A Cross-Architecture Evaluation

Young, Richard J.

A cross-architecture comparison of abliteration techniques used to remove refusal behavior from open-weight LLMs, examining trade-offs in capability retention, residual safety, and how each method generalizes across model families.

Replicating TEMPEST at Scale: Multi-Turn Adversarial Attacks Against Trillion-Parameter Frontier Models

Young, Richard J.

This study evaluates large language model vulnerability to sophisticated multi-turn attacks using the TEMPEST framework across ten frontier models from eight vendors, testing 1,000 harmful behaviors with over 97,000 API queries. Six models showed 96-100% attack success rate (ASR), while four demonstrated greater resilience with ASR between 42-78%. Enabling extended reasoning reduced ASR from 97% to 42%. The research concludes that safety alignment quality varies substantially across vendors, model scale does not ensure adversarial robustness, and deliberative inference represents a promising defense approach. Current alignment techniques remain fundamentally vulnerable to adaptive multi-turn attacks regardless of model scale, though thinking modes offer potential mitigation strategies.

Evaluating the Robustness of Large Language Model Safety Guardrails Against Adversarial Attacks

Young, Richard J.

This research examined ten publicly accessible guardrail models from major organizations including Meta, Google, IBM, NVIDIA, Alibaba, and Allen AI, testing them against 1,445 prompts across 21 attack categories. While Qwen3Guard-8B demonstrated the strongest overall accuracy at 85.3%, all models substantially underperformed when encountering novel attacks compared to benchmark prompts. Qwen3Guard dropped from 91.0% to 33.8% (a 57.2 percentage point gap), whereas Granite-Guardian-3.2-5B exhibited superior generalization with only a 6.5% performance decline. The research also identified a previously unknown failure mode where certain guardrail models generated harmful responses when prompted with a helpful mode jailbreak technique. Generalization capacity, rather than benchmark accuracy, represents the more meaningful evaluation criterion for assessing safety guardrail effectiveness.

Can Large Language Models Address Problem Gambling? Expert Insights from Gambling Treatment Professionals

Ghaharian, Kasra, Soligo, Maximilliano, Young, Richard, Golab, Lukasz, Kraus, Shane W., Wells, Samantha

Co-authored peer-reviewed work with gambling treatment professionals evaluating LLM responses to problem gambling questions, aimed at ensuring LLM-based education and support tools reduce harm rather than inadvertently encouraging risky behavior.

Evaluating LLM Responses to Problem Gambling Vignettes

Ghaharian, Kasra, Kraus, Shane W., Young, Richard, Golab, Lukasz, Wells, Samantha, Soligo, Maximilliano

A preprint evaluating LLM responses to a set of problem-gambling clinical vignettes, assessing harm avoidance, accuracy, and alignment with treatment-professional expectations.

Clinical and Cost Outcomes in Medicare Advantage Following Post-Acute Care

Gada, Eshan, Pangburn, Paula, Sahr, Christopher, Jarvis, Mary S., Young, Richard J.

An analysis of clinical and cost outcomes for Medicare Advantage members following discharge to post-acute care, identifying patterns that predict readmission and total cost of care.

Comparative Analysis of LoRA-Adapted Embedding Models for Clinical Cardiology Text Representation

Young, Richard J., Matthews, Alice M.

This study evaluated ten transformer-based embedding models fine-tuned for cardiology using Low-Rank Adaptation on over 106,000 cardiology text pairs from medical textbooks. Encoder-only architectures, particularly BioLinkBERT, achieve superior domain-specific performance (separation score: 0.510) while using fewer computational resources than larger models. The study challenges the notion that bigger language models automatically produce better clinical embeddings and offers practical recommendations for developing medical NLP systems. All models, code, and datasets are publicly available.

Vision Token Masking Alone Cannot Prevent PHI Leakage in Medical Document OCR: A Systematic Evaluation

Young, Richard J.

This research examines how vision-language models used for document scanning in healthcare handle sensitive patient information. Testing seven different masking approaches on DeepSeek-OCR using 100 synthetic medical billing statements, all strategies achieve only 42.9% protection against health information exposure. While long-form identifiers like patient names and addresses are successfully blocked, shorter structured data such as medical record and social security numbers leak through. This is attributed to the language model's contextual reasoning rather than inadequate visual masking. Combining vision masking with language-level processing reaches 88.6% effectiveness, recommending future work focus on decoder-level adjustments for HIPAA compliance in medical document handling.

CardioEmbed: Domain-Specialized Text Embeddings for Clinical Cardiology

Young, Richard J., Matthews, Alice M.

Benchmarking Multiple Large Language Models for Automated Clinical Trial Data Extraction in Aging Research

Young, Richard J., Matthews, Alice, Poston, Brach

A benchmark of five LLMs and a multi-LLM ensemble for extracting structured protocol fields from ClinicalTrials.gov, applied to aging-related tDCS trials. The pipeline doubled eligible-trial retrieval to 83 studies and reached near-perfect agreement on core stimulation parameters (kappa approximately 0.94; ICC approximately 0.95-0.96), reducing manual screening burden.

Across the Bettor-Verse: An Open Banking Perspective on Gambling in the United Kingdom

Ghaharian, Kasra, Peterson, Jay, Young, Richard

An open-banking analysis of gambling transaction patterns among UK consumers, characterizing behaviors associated with progression from recreational to problem gambling using real transaction-level data.

Beyond the Bet: An Open Banking Perspective on Gambling in the UK

Ghaharian, Kasra, Peterson, Jay, Young, Richard

Preprint version of the open-banking gambling analysis later published in Journal of Gambling Studies, characterizing gambling behavior patterns in UK transaction data.

UNLV International Gaming Institute - Open Banking Data Research

Ghaharian, Kasra, Azizsoltani, Mohammad, Puranik, Pranav A., Young, Richard

A research dataset and methodology report documenting the UNLV IGI open-banking gambling research program, including data structure, governance, and analytic approach.

Unfulfilled Home Health Referrals Lead to Higher Mortality Among Medicare Advantage Members

Gada, Eshan, Pangburn, Paula, Sahr, Christopher, Schaben, Carol, Young, Richard

An analysis of Medicare Advantage members showing that unfulfilled home-health referrals are associated with significantly higher mortality, highlighting a measurable care-coordination gap with patient-safety implications.

Development of an Artificial Intelligence Approach That Employs Genomic and Brain Imaging Features to Improve the Diagnosis of Alzheimer's Disease (AD)

Zhuang, Xiaowei, Young, Richard, Tillett, Ryan, Cordes, Dietmar, Oh, Edwin

A multimodal AI approach combining genomic features with brain imaging biomarkers to improve diagnostic accuracy for Alzheimer's disease relative to single-modality baselines.

When Models Can't Follow: Testing Instruction Adherence Across 256 LLMs

Young, Richard J., Gillins, Brandon, Matthews, Alice M.

Despite widespread deployment of Large Language Models, systematic evaluation of instruction-following capabilities remains challenging. While comprehensive benchmarks exist, focused assessments that quickly diagnose specific instruction adherence patterns are valuable. As newer models may be trained on existing benchmarks, novel evaluation approaches are needed to assess genuine capabilities rather than memorized performance. This paper presents a streamlined evaluation framework using twenty carefully designed prompts to assess LLM instruction-following across diverse task categories. We demonstrate this framework through a large-scale empirical study conducted on October 14, 2025, testing 256 verified working models from 331 available via OpenRouter. To ensure methodological rigor and prevent selection bias, we first verified each model's basic functionality before inclusion. Unlike large-scale benchmarks requiring extensive computational resources, our approach offers a practical diagnostic tool researchers and practitioners can readily apply. Our methodology builds upon verifiable instructions while introducing a compact test suite balancing comprehensiveness with efficiency. Each prompt targets distinct aspects of instruction following, including format compliance, content constraints, logical sequencing, and multi-step task execution. We evaluate models from major providers (OpenAI, Anthropic, Google, Meta, Mistral) and emerging implementations (Qwen, DeepSeek, community models), providing comparative performance analysis. Our findings reveal consistent failure modes and identify specific instruction types posing particular challenges. This work contributes both a practical evaluation tool and one of the most comprehensive empirical analyses of instruction-following capabilities across the contemporary LLM landscape.

The Influence of Transcranial Alternating Current Stimulation on the Excitability of the Unstimulated Contralateral Primary Motor Cortex

Wilkins, Erik W., Young, Richard J., Davidson, Ryan, Krider, Robert, Alhwayek, Ghaith, Park, Jin A., Parikh, Arjun C., Riley, Zachary A., Poston, Brach

An investigation of how transcranial alternating current stimulation applied unilaterally influences excitability in the unstimulated contralateral primary motor cortex, with implications for bilateral neuromodulation protocols.

Non-Dominant Hemisphere Excitability Is Unaffected During and After Transcranial Direct Current Stimulation of the Dominant Hemisphere

Wilkins, Erik W., Young, Richard J., Houston, Daniel, Kawana, Eric, Mora, Eduardo, Sunkara, Manish, Riley, Zachary A., Poston, Brach

This study tests whether transcranial direct current stimulation of the dominant hemisphere produces measurable changes in the non-dominant hemisphere, finding no significant cross-hemispheric effects on excitability during or after stimulation.

The Influence of Transcranial Alternating Current Stimulation on Fatigue Resistance

De Guzman, Kelly A., Young, Richard J., Contini, Vincent, Clinton, Ethan, Hitchcock, Amelia, Riley, Zachary A., Poston, Brach

This study investigates how transcranial alternating current stimulation applied to motor cortex influences fatigue resistance during sustained voluntary muscle contractions.

Minimotifs Dysfunction Is Pervasive in Neurodegenerative Disorders

Sharma, Sandeep, Young, Richard J., Chen, Jingchun, Chen, Xuemei, Oh, Edwin C., Schiller, Martin R.

Analysis showing that short functional protein motifs (minimotifs) are systematically disrupted across neurodegenerative disorders, suggesting a shared molecular vulnerability and a target class for therapeutic study.

Minimotif Miner 4: A Million Peptide Minimotifs and Counting

Lyon, Kenneth F., Cai, Xinge, Young, Richard J., Mamun, Abdullah-Al, Rajasekaran, Sanguthevar, Schiller, Martin R.

The fourth release of the Minimotif Miner database, expanding annotated peptide minimotifs to the million-entry scale and adding new search and visualization tooling for functional motif discovery.

Natural Variability of Minimotifs in 1092 People Indicates That Minimotifs Are Targets of Evolution

Lyon, Kenneth F., Strong, Christy L., Schooler, Sara G., Young, Richard J., Roy, Nikhil, Ozar, Bill, Bachmeier, Michael, Rajasekaran, Sanguthevar, Schiller, Martin R.

Analysis of 1,092 human genomes showing that short functional protein motifs (minimotifs) display patterns of natural variability consistent with being targets of evolutionary selection.