A Self-Discovering Abstraction Engine via Variational Symmetry, Typed Program Induction, and Neural-Symbolic Compilation
author : Tofara Moyo
abstract
We present a general-purpose abstraction engine capable of discovering, compressing, and reusing structure across problem instances through a self-referential learning loop. The system integrates variational inference over program partitions, category-theoretic symmetry composition , typed abstraction languages, neural-symbolic program compilation, and curriculum-driven abstraction growth. Unlike task-specific solvers, the proposed architecture treats programs themselves as objects of inference, enabling recursive abstraction, symmetry quotienting, and compression-driven primitive discovery. While motivated by the Abstraction and Reasoning Corpus (ARC), the framework is domain-agnostic and targets the broader problem of learning abstractions that generalize across tasks.