People
Join the group
We are looking for curious, motivated researchers to join the group for their PhD. If you're interested in the intersection of language, AI, and the brain — have a look at our open positions
How do we understand language? What does the brain actually do when we read or hear a sentence? And how is it that Large Language Models — trained on nothing but predicting the next word — have not only mastered language, but also turned out to be the most accurate models of human brain responses to language?
In the LPC Group, we use the tools of modern AI to model language in the human mind and brain. We also run the logic in reverse: drawing on what psychology and neuroscience teach us about human language processing, we build language models constrained by the human cognitive architecture.
Ultimately, we aim to understand how the human brain learns and represents language, and to build more cognitively faithful models of human language processing.
We are looking for curious, motivated researchers to join the group for their PhD. If you're interested in the intersection of language, AI, and the brain — have a look at our open positions
Our group works at the intersection of cognitive science and AI, trying to understand how brains and minds process language and make sense of the world. A recurring idea is that prediction is a core principle of both biological and artificial intelligence. Language is our primary focus, but we also venture into other domains - how we see, reason, and remember. More on specific research directions below.
As you read or listen, your brain is constantly predicting what comes next, much like a language model. But what exactly is the brain predicting? Words? Syntax? Meaning? And is human language comprehension merely a probabilistic guessing game — like in AI models — or is there more going on? We use language models combined with eye tracking and neuroimaging (fMRI, MEG) to find out.
LLMs are powerful but cognitively implausible — they have superhuman memory, start as a blank slate, and train on more text than any human encounters in a lifetime. Can we do better? We try to build models that learn from realistic data and generalise more like children do. We then test these models and compare them to human language processing. The goal is to build more cognitively faithful models, and use neural networks as a testbed for theories of human language learning.
Language models have become our best tools for predicting how the brain responds to language, but we don't really know why. What is it about these models that makes their representations so apparently brain-like? We use interpretability methods from AI to try and find out. The goal is to understand which aspects of language processing are shared between artificial and biological systems — and which are not.
Prediction is not unique to language. We study how the brain builds predictive models of the visual world — from simple sequences to intuitive physics — and ask whether the intelligence emerging in large-scale autoregressive models extends beyond language to general reasoning. This thread connects our core interest in prediction to broader questions about the nature of intelligence.