Multi-Agent RAG for Evidence-Based Counterspeech →
Specialized agents retrieve from static and dynamic sources, summarize evidence, generate a response, and refine it for factuality and relevance.
Research highlights
Trustworthy language systems, from evidence to impact.
My research asks how language models can find reliable evidence, communicate it at the right level, and remain dependable in high-stakes health and crisis settings.
Grounding & knowledge
Evidence-grounded language models
I develop retrieval-augmented, knowledge-graph, and multi-agent methods that connect model responses to authoritative and timely sources. The goal is not simply to retrieve more text, but to select the evidence needed for a specific claim and make the resulting answer easier to verify.
- Retrieval-Augmented Generation
- Knowledge Graphs
- Multi-Agent Systems
- Evidence Attribution
Questions I study
- How should multiple agents retrieve, synthesize, and critique evidence?
- How can structured knowledge expose missing context and latent information needs?
- Which evidence is sufficient to support a factual, traceable response?
Selected projects
Published workSourceMinds: NLI-Grounded Citation Auditing ↗
A multi-agent fact-checking pipeline that audits whether generated claims are actually supported by their cited sources.
SciClaimSeekers: Scientific Source Retrieval ↗
A scientific-source retrieval system that combines broad candidate retrieval with LLM-based reranking for social-media claims.
People & interaction
Adaptive human–AI interaction
I design systems that adapt explanations to a reader’s literacy, knowledge, intent, and uncertainty. This work treats communication as an interaction: a model may need to adjust explanatory depth, ask a focused question, or answer directly while preserving the same factual foundation.
- Controllable Generation
- Health Literacy
- Personalization
- Proactive Clarification
Questions I study
- How can one evidence base support explanations at different levels of detail?
- When should an AI ask for clarification instead of answering immediately?
- How can adaptation improve usefulness without changing factual meaning?
Selected projects
Published & ongoing workSpeaking at the Right Level →
A RAG and reinforcement-learning framework that generates grounded corrections for readers with different health-literacy needs.
Ask or Answer: Multi-Turn Intervention ↗
A decision framework that learns whether to ask for missing information or provide a direct health-misinformation intervention.
Granularity-Aligned Health-Misinformation Refutation →
A dual-path knowledge-graph approach for matching biomedical causal depth and explanatory detail to the intended reader.
Measurement & reliability
Evaluation for high-stakes NLP
I build benchmarks, metrics, and human-evaluation protocols that test whether a system is factual, consistent, useful, and sensitive to context. Health and crisis communication provide demanding test beds because a fluent response can still fail through stale evidence, misplaced information, or inappropriate detail.
- LLM Evaluation
- Diagnostic Benchmarks
- Human Evaluation
- Crisis Informatics
Questions I study
- Does a response use evidence from the right place and time?
- Can automated judges measure factuality and audience alignment reliably?
- Which metrics reflect what readers actually find understandable and useful?
Selected projects
Published workA Dynamic Fusion Model for Consistent Crisis Response ↗
A framework and evaluation metric for producing consistent, professional, and actionable responses during crises.
Dynamic Fusion of LLMs for Crisis Communication ↗
A retrieval-supported approach that combines complementary language models for timely, actionable crisis communication.
Can LLMs Reason Over Where and When? Benchmarking Spatiotemporal Synthesis in Crisis Response ↗
Evidence-coverage auditing and synthetic report enrichment support ST-CrisisSynth’s query-dependent semantic, spatial, and temporal re-ranking. Human evaluation finds a 70.0% correct-response rate, 3.3 percentage points above the strongest baseline.
Ongoing research
Under reviewSpatiotemporal Reasoning for Crisis Response
An ongoing study of how language models retrieve and reason over location- and time-sensitive information during crises.
A connected agenda
Ground the answer. Adapt the explanation. Test what matters.
These areas reinforce one another: reliable evidence supports adaptation, human needs define meaningful evaluation, and rigorous evaluation reveals where grounding and interaction methods must improve.
See the complete publication record, including papers beyond these selected projects.
View all publications