Meta-Generation: A General Framework for Generating Generators, Generative Mechanisms, and the Rules That Construct Them

Generative artificial intelligence has primarily focused on generating content, including text, images, audio, video, and executable programs. A broader possibility is to generate not only outputs but also the mechanisms responsible for producing those outputs. This possibility raises a fundamental question: can a generative system construct a generator, create the procedure that constructs generators, and generate or modify the rules governing that construction process? This paper introduces Meta-Generation, a general framework for investigating the generation of generative mechanisms as an explicit computational objective. The framework distinguishes content generation, generator construction, generative-mechanism construction, and the construction or modification of procedures that govern generative-mechanism construction. It represents these operations as typed transformations over generators, programs, computational graphs, and construction procedures. A cost-aware optimization objective is proposed to evaluate generated mechanisms according to task performance, computational expenditure, complexity, robustness, and validity. The framework also introduces a mechanism archive, candidate selection, controlled modification, and optional recursive generation as components of an implementable system. Importantly, Meta-Generation is not presented as an entirely unprecedented concept. Hypernetworks, neural architecture search, meta-learning, program synthesis, and recursive self-improvement already address important parts of this problem. Recent work on HyperAgents, in particular, explores the modification of procedures responsible for generating future improvements. The proposed contribution is a general formalization that treats multiple kinds of generative construction as a unified research problem and identifies conditions under which this formulation might provide practical advantages. Testable hypotheses and an experimental methodology are presented. No empirical superiority is claimed. The central proposition is that the methods by which generative systems are constructed may themselves become objects of systematic generation, evaluation, and optimization.

Keywords: Meta-Generation; generative models; generator con