About his postdoctoral project

Large language models (LLMs) trained on naturalistic data learn high-dimensional representations that support flexible, context-dependent inference. In parallel, neural activity underlying human decision-making exhibits complex, distributed dynamics that remain difficult to characterize using traditional experimental approaches. Although LLMs are trained on linguistic data, their emergent behavior suggests that their internal representations may capture abstract computational structures that are also relevant to decision-making processes.

This project investigates whether internal representations of large language models align with human neural activity during decision-making. We extract representations from pretrained models and evaluate their ability to predict neural signals. By comparing representations across model architectures and layers, we aim to characterize representational features that may correspond to decision-related neural dynamics.