Abstract
Current LLMs are able to solve incredibly complex tasks, but these are very
hard to interpret. Our goal is to develop a suite of tasks that are minimalist but complex enough to
extract complex behavior dynamics and latent cognitive computations.