My current work with STAIR the lab explores LLMs in strategic settings. It is interested in whether LLM agents leverage concrete knowledge from pre-training or attempt to “reason from scratch” when facing novel environments. Initially modeled by their actions in 2-player multi-turn Matrix games.
In prior work with the lab [@UIUC 2017-2020] I worked on several projects on inductive biases for generative models. Including model regularizations to induce fairness properties and latent space disentanglement.