Research | Xinyi Liu

I study how language models understand complex human situations, whether their judgments remain consistent in action, and how AI systems can better support human agency over time.

My research connects natural language processing, AI evaluation, and human-centered AI. The questions below guide my current work and longer-term research agenda.

01 / Judgment → action

LLM–Agent Consistency

Do models preserve their judgments as they move from understanding a situation to responding and acting?

Conceptual illustration

Research in development

From judgments to agent actions

I am developing an evaluation framework for consistency across static diagnosis, response generation, and simulated tool use. A central question is whether a model that recognizes a conflict with human wellbeing or agency preserves that judgment when asked to act.

02 / Language → structure

Structured Language Understanding

Can language models recover complex relations and their interdependencies from natural language?

Conceptual illustration

Ongoing research

Recovering interdependent relations from text

My current work investigates how language models reconstruct coupled relations in natural-language accounts of user experiences. I focus on dependencies between relations and cases where the same relation instance participates in multiple structures, testing what models preserve, miss, or connect incorrectly.

03 / AI systems ↔ human lives

Human-Centered AI Alignment

How can AI systems support human goals, wellbeing, and agency over time?

Conceptual illustration

Research agenda

AI, wellbeing, and human agency

I explore how AI-shaped environments interact with human needs and capacities over time. This agenda asks how systems can reduce persistent pressures on wellbeing while preserving people's ability to reflect, choose, and pursue their own goals.

Foundations / Visualization & human–AI interaction