Our mission is to enable trust between humans and AI agents through language technology, making human-machine teaming possible in real-world applications. Our research centers on natural language processing, with broader contributions to machine learning, multimodal models, and artificial intelligence.
We aim to develop an agentic AI system that automates key aspects of mathematical reasoning, including theorem decomposition, lemma identification, novel proof strategy generation, seamless translation between natural language and formal proof systems, and automatic conjecture discovery.
We develop multimodal systems that learn new concepts through language, visual grounding, interaction, and feedback, with emphasis on recognizing unfamiliar objects, understanding richer descriptions, robust video-language understanding, and more reliable human-machine collaboration.
We integrate neural prediction with probabilistic inference and symbolic constraints so models can produce structured outputs that are accurate and verifiable. We apply these approaches to problems ranging from improving LLM safety to enhancing model reasoning capabilities.
We build detection, evaluation, guardrail, and mitigation methods for unsafe behavior, over-refusal, sleeper agents, jailbreaks, and customized multimodal safety policies.
We develop open-source language agents, tool-use workflows, long-term memory benchmarks, and data-analysis agents for complex mulitmodal interactive tasks.
We develop trustworthy AI solutions for healthcare applications, from matching patients in clinical trails, to clinical report analysis, radiology summarization, and patient-centered medical decision support.
We develop NLP approaches to detect early signs of emerging infectious diseases, predict their spread, and detect and monitor risk factors through multilingual social media posts.
We study how bias appears in representations, generation, recommendations, and social text, and design human-centered interventions for more equitable AI systems.
This project studies commonsense knowledge from video, images, text, and knowledge bases, with benchmarks and models for multimodal social and scientific reasoning.