Research

At LINK Lab, we study how language and artificial intelligence systems understand, reason about, and interact within human and social contexts.

Our research spans the evaluation and development of large language models, with a particular focus on socially grounded reasoning, human-centered applications, domain-specific knowledge, and the safety and reliability of AI systems in real-world settings.

• Socially Grounded Language Models 

We study how language models understand and respond to social contexts, including personality, social roles, norms, preferences, and interpersonal relationships.

Our goal is to develop methods for evaluating whether language models behave consistently and appropriately across diverse social situations.

• Mental Health and Domain-Specific AI 

We develop and evaluate language models for specialized domains where accurate knowledge and careful reasoning are essential.

Our research focuses on integrating domain knowledge into language models and evaluating how models reason, acquire information, and make decisions in applications such as mental health.

• Agentic AI Systems  

We study AI agents that can reason, plan, interact, and make decisions over extended sequences of actions.

Our research investigates how agentic AI systems behave in dynamic and socially complex environments, including how they adapt to context, coordinate with people and other agents, and make decisions under competing goals and constraints.

• AI Safety 

We study the behaviors and failure modes of LLMs and agentic AI systems to support the development of safer and more reliable AI.

Our research focuses on identifying and evaluating issues such as bias, inconsistent behavior, misalignment, unsafe decision-making, and failures that emerge when AI systems interact with people or operate in complex environments.