Our work investigates the intersection of recursive self-improvement, automated theorem proving, and program synthesis within the Gödel machine framework
We study self-referential systems capable of modifying their own learning algorithms. Our current work explores how a Gödel machine — as formalized by Schmidhuber (2006) — can leverage automated proof search to identify and apply modifications to its own codebase when a proof of expected utility improvement is found. We are investigating tractable approximations to the proof search problem and their applications to neural architecture optimization and policy improvement.