Research Program

How I came to these questions

My undergraduate studies in philosophy drew me to logic: a way to make the structure of an argument explicit and examine what follows from it. In my thesis, I studied Leibniz’s logical system and his proposal for a universal symbolic language. This raised a question that continues to interest me: how do formal descriptions of meaning relate to the ways people actually understand and reason?

Learning languages gave that question an everyday setting. I found myself paying attention to how sentence structure contributes to meaning, and how different languages can express similar ideas through different structures. This led me to ask when differences in linguistic form affect understanding and thought. Alongside philosophy, I studied psychology and took part in a project on children’s learning difficulties. My master’s studies in Human and Artificial Intelligence have brought these interests into experimental and computational work.

From a question about people to an experiment on a model

Linguistics and cognitive science provide precise questions about meaning and decades of carefully measured human behaviour. Taken over to language models, those questions become controlled experiments: tests of behaviour show what a model does, and interpretability methods show where in the model it happens. The work proceeds in four steps.

1. A question with a precise contrast

The question comes with alternatives that predict different outcomes. When a modifier gives a dog wings, does a model update the properties that depend on the new feature, or those that merely co-occur with it? When moral wording changes a model’s dilemma choice, has the weighing of outcomes changed, or a tendency to answer with an action? My writing in semantics and pragmatics, including typicality in referent choice, is where distinctions of this kind are worked out.

2. What the human data measure

Human studies show what a task measures and what it leaves open. My master’s thesis on modal language in preschool children and a course experiment on fact-checking taught me how such data are produced; a child’s answer does not, by itself, settle what concepts the child has. In the modifier study, human property and association norms supply predictors fixed outside the models.

3. Behaviour under matched controls

The same contrast is then posed to language models on matched items: bare nouns, common modifiers, and modifiers that are just as unusual but unrelated to the property asked about; in the moral-framing project, checks of answer labels, option positions and task understanding.

4. Inside the model

Interventions on internal representations test where a behaviour arises: residual-stream patching and low-rank analysis in the moral-framing project, and an exploratory patching analysis in the modifier study.

Human and model studies can inform each other without assuming that they use the same mechanisms. Similar answers can arise from different processes, and an intervention in a model does not establish a mechanism in people.