Learning History-Aware Paraphrase Robustness in LLMs under Epistemic Uncertainty from Dependent Logs
Robust LLM training under epistemic uncertainty, distribution shift, and dependent conversational logs.
LLMsSFTUncertainty
PUBLICATIONS & PROJECTS
Publications and projects.
Current AI work and earlier statistical research, with project details and available papers, slides, and code.
Neyman Pearson Classification for Heart Transplant Rejection Accounting for Model Selection Uncertainty
Clinical classification with controlled false positive rates, stable biomarker selection, and repeated sample splitting.
HealthcareClassificationUncertainty
Duplicates in Prior Authorization Data: Prevalence, Implications, and Strategies for Mitigating Privacy Risks
Collaborative work on privacy risks and data-quality issues in prior authorization records.
PrivacyData QualityHealthcare
In process
LLM Agent for Crash Analytics and Adaptive Feature Enrichment
An agent workflow that enriches structured crash records with external context and produces analysis-ready case reports.
LLM AgentsTransportationApplied AI
Uncertainty Assessment for Paraphrase Robust LLMs
Credible robustness assessment for history-aware LLM behavior under paraphrase variation and dependent evaluation logs.
LLM EvaluationRobustnessUncertainty
Private Minimum Hellinger Distance Estimation via Hellinger Distance Differential Privacy
Privacy-aware estimation with robust inferential guarantees and Hellinger-based privacy machinery.
Differential PrivacyRobust Inference
Sharp Large Deviations and Gibbs Conditioning for Threshold Models in Portfolio Credit Risk
Sharp asymptotics, Value at Risk, Expected Shortfall, and Gibbs conditioning for rare portfolio-loss events.
Credit RiskLarge Deviations
Local Limit Theorem and Large Deviations for Branching Process with Immigration
Local limits, oscillation behavior, and large deviations across branching-process growth regimes.
Branching ProcessesLarge Deviations
Private and Robust Inference for Hawkes Process Data and Integer-Valued Autoregression
Robust and privacy-aware inference for event data, linking Hawkes processes and integer-valued autoregression.
Hawkes ProcessesPrivacyInference