Laying the foundations for foundation models of cognitive science

Foundation models, large-scale artificial intelligence (AI) systems trained on vast amounts of data and able to engage with a wide range of tasks, have captivated public attention and promise transformative impact across many fields, from mathematics to education (Bommasani et al, 2021). The capacity for such models to fluidly engage in natural language carries huge potential implications for cognitive science: for the first time, we have computational models that bring us a significant step toward the generality of human cognition, thereby inviting us to reconceptualize how we may build models of the human mind (Binz et al, 2025; Wong & Collins et al, 2025). Yet, major questions remain as to how to build such models, what kind of data is needed for this, and what the implications are for cognitive science.

This workshop brings together researchers to examine how foundation models can inform and advance cognitive modeling, and what it may take to build towards a grand goal of general-purpose foundation models of human cognition. To this end, we have invited leading experts to present recent advances across a broad set of questions, including:

how to combine black-box models with existing theories about human cognition (Griffiths et al, 2025)
how to use foundation models as part of a broader toolkit for building cognitive models (Wong & Collins et al, 2025, Wust et al, 2025);
how to integrate language into cognitive architectures (Wray et al, 2025);
what would it mean to have a true foundation model of human cognition that can capture both population-level regularities and individual differences;
how do we need to evaluate such models to obtain insights that human scientists care about?

In doing so, we hope to foster a dialogue on how to incorporate key characteristics of foundation models into cognitive modeling. This powerful synthesis -- of advances in large-scale language modeling, alongside other computational approaches developed in computational cognitive science (e.g., for program synthesis) -- could soon lead to computational models that capture human behavior not just within a single experiment or domain, but across many of them, taking steps towards a long-standing goal of cognitive science (Newell, 1990). This offers a promising route toward addressing one of psychology's biggest challenges -- the generalization crisis -- by enabling models to synthesize findings across traditionally isolated areas of cognition. Furthermore, we envision that such models will be stimulus-computable, able to operate on rich, naturalistic stimuli, such as raw text or images, thereby moving cognitive modeling toward more ecologically valid environments.

Goals and scope

The goal of this workshop is to lay the foundations for foundation models of cognition -- models that are capable of reasoning about any kind of problem people can, and ideally do so in a way that advances our understanding of human cognition. Building such models raises a number of fundamental questions, including: (1) how can and should such models be built? (2) at what level of abstraction should they operate? (3) what are evaluation protocols that allow us to meaningfully compare competing models against each other and against humans? (4) what are the implications for cognitive science as a field if we can build such models? By tackling these questions, the workshop will serve as a catalyst for new research directions. We expect it to kickstart collaborations, inspire new modeling approaches, datasets, and evaluations for such models, and help chart a roadmap for years (even decades) to come.