Research

Using Messy Text in Future LLM Annotations

Researchers often inherit annotations built over years of costly human coding, but noise and ambiguity in older corpora make them unusable for modern text-as-data pipelines. This research is supported by an NSF grant.

• Deployed a stateful and observable orchestration layer for LLM workflows on HPC with batch processing to realign annotations from noisy corpora with a codebook, producing structured, auditable, and source-grounded training data for downstream fine-tuning.
• Reduces weeks of manual labor to hours of parallelized computation while retaining span-level evidence.

Do Newer Vision-Language Models Detect Better? Evaluating UXO Detection Performance on a High-Risk UAV Hyperspectral Landmine Dataset

• Designed controlled perturbation experiments to benchmark multiple generations of vision-language models on hyperspectral anomaly detection, identifying systematic model weaknesses and quantifying robustness and performance degradation under input variation.

How AI Is Governed: LLM-driven Entity and Relation Extraction on State AI Bills

Deployed an agentic RAG pipeline for named entity recognition and relation extraction on state AI legislation, fine-tuned a base model with LoRA to predict policy adoption, and evaluated performance using cross-validation, multi-seed testing, calibration analysis, bias-aware metrics, and compute-performance trade-offs.

Contextualized Synthetic Control: Assessing the Impact of the NVIDIA Chip Ban on the AI Arms Race

Extended the synthetic control method with a custom multi-head attention mechanism to construct counterfactuals for treated units with limited comparable donors, enabling GPU-accelerated causal inference under sparse and highly dissimilar donor pools.

The Scope of Social Science Research Tasks Using LLMs

Designed a benchmark to quantify systematic LLM biases across research tasks and high-risk corpora, identifying model limitations and developing evidence-based guidance for responsible LLM use in research.

• Measures bias across task and corpus risk levels rather than relying on generic QA and multiple-choice benchmarks.
• Established reproducible task-specific best practice guidelines for researchers, replacing trial-and-error with task-specific thresholds.

LLM-Based Rhetoric Extraction from Censorship-Related Legal Texts

This paper extracts rights and security related rhetoric from cross-national censorship legislation and converted unstructured corpora into structured dataset.

• Built LLM extraction pipelines on a cross-national censorship law corpus, producing country-year measures of content control language with traceable verbatim span evidence for comparative censorship law research at scale.
• Benchmarks two information extraction pipelines, establishing an auditable method for measuring censorship rhetoric across legal corpora.