Discovering and Engineering the Computation Underlying Large Intelligent Agents

Abstract

The richness of language and intelligent behavior has often been attributed to latent compositional structure. Can we build tools for discovering how deep networks learn and represent this latent structure implicitly? And more importantly, can we use this knowledge to improve generalization in largely structure-less general purpose models or refine our understanding of the world they describe? In this dissertation, I present three perspectives to answer these questions. First, I present experimental methods to functionally characterize the space of learnt solutions in LLMs and demonstrate how this understanding can be used to improve their empirical generalization in a gradient free manner, sometimes by as much as 30% points on language understanding benchmarks. Following that, I show how to decipher the structure of another (black box) language-like system, the naturally arising communication system of sperm whales in the wild, discovering for the first time a unique combinatorial communication system. Finally, I apply insights from these results to equip embodied agents with a latent language of thought—hierarchical and compositional—and show how it can enable long-horizon reasoning and planning in these systems. This dissertation ultimately aims to bridge the gap between natural and artificial intelligence, offering new insights into both the fundamental nature of communication in complex biological organisms in the wild and the development of more powerful, and improved AI systems. A key pattern in the discoveries in this thesis has been how simple structures enable complex externalized behaviors in both biological organisms and AI systems.