Research

I study how experts think through long, difficult tasks, and build AI that helps them think rather than doing the thinking for them.

Language models write fluently, but they do not work the way experts do. When we compared lawyers' recorded workflows with an LLM's on the same legal task, the lawyers followed the one branch that mattered and revised their plan as the case developed; the model expanded every branch it could find.

Two workflow graphs side by side. The human lawyer follows one relevant branch; the LLM expands every possible branch. — Human vs. LLM workflow on the same legal task. From LawFlow (COLM 2025), with UMN Law School and Thomson Reuters.

Most models learn only from finished text. We also record the process behind it, then use those traces to train models, test them against people, and build tools that fit how experts already work.

Longer versions: research statement (non-technical, technical) and my PhD dissertation. Every paper is on the Publications page.

The big picture

Three connected pillars. Data and models from the first feed the tools in the second; the third asks whose perspectives all of it represents. Pick a topic to see the projects behind it.

Cognitive scaffolding

Can models learn from how people actually think?

Writing a paper takes months of planning, drafting, abandoning, and revising, yet a model trained on the final PDF sees none of that. We record that process, use it as training signal, and borrow methods from cognitive science to measure where models still differ from people.

Left: the scientific research cycle. Right: one project's stages, intents, sub-tasks, and tools, from our collected workflow data. — One research project traced from discovery to writing: stages, intents, sub-tasks, and the tools used at each step.

Thinking assistants for experts

What does an AI that helps experts think look like?

In science and law, full automation tends to flatten the reasoning that makes expert work valuable. We build tools that help at the moment of difficulty, adapt to how a particular expert works, and leave the judgment with the person.

Left: an augmented PDF reader explaining terms in an equation. Right: an in-situ writing assistant flagging an overgeneralized citation claim. — Augmented reading (left) and proactive, in-situ writing feedback (right).

Societal alignment

Whose views does a model represent?

People disagree, and their perspectives differ by background, identity, and culture; a model trained toward an average answer can erase that. I founded the Pluralistic Alignment workshop (NeurIPS 2024; second edition at ICML 2026) to work on this with ML, HCI, and social-science researchers.

Collaborators and communities

The work draws on linguistics, cognitive science, and the social sciences, with collaborators in computer science, law, psychology, education, journalism, design, and medicine. To connect these communities I co-organized the first CtrlGen workshop on controllable generative modeling (NeurIPS 2021), founded the In2Writing workshop series on intelligent and interactive writing assistants (ACL 2022, CHI 2023, CHI 2024), and founded the Pluralistic Alignment workshop (NeurIPS 2024).