Research
My research is in natural language processing and large language models, focused on making LLMs trustworthy and human-centered for high-stakes communication. I work across three connected directions: knowledge-grounded generation, controllable and personalized generation, and rigorous evaluation, with applications in health misinformation and crisis response.
Knowledge-Grounded & Retrieval-Augmented Generation
I build retrieval-augmented generation (RAG) and knowledge-graph methods that ground LLM outputs in trusted external evidence, combining static authoritative sources with dynamic retrieval to improve factual reliability and reduce unsupported claims.
My work explores multi-agent orchestration, evidence selection, and structured knowledge for robust long-form generation.
Controllable & Personalized Generation
I research generation methods that adapt LLM responses to individual users, including health literacy, comprehension, and readability constraints, so that outputs are both accurate and appropriate for their audience.
This includes literacy-controlled generation and multi-turn decision frameworks that learn when to ask a clarifying question versus when to respond.
LLM Evaluation & Benchmarking
I design rigorous evaluation frameworks, benchmarks, and quality metrics to assess LLM behavior, factuality, and reliability, using both automated methods and human evaluation.
My work includes diagnostic benchmarks for probing model reasoning and evaluation pipelines that measure faithfulness and response quality.
Applications: Crisis & Health Communication
I apply these methods to high-stakes domains. In crisis communication, I study consistent, actionable response generation, dynamic fusion across models, and spatiotemporal reasoning over situational-awareness information.
In public health, I work on evidence-based counterspeech and health misinformation mitigation that supports safer communication without sacrificing factual precision.