Latent Meta-Abstraction over Transformer Representations: A Two-System Architecture for Persistent Meta-Abstraction and Compute Leverage

Abstract

This concept paper proposes a two-system architecture pairing a conventional transformer (Primary Model) with a separately parameterized Meta-Abstraction Model restricted operationally to activation-derived Primary representations and its own latent memory. A smaller or lower-cost specialized learner could accumulate activation-derived experience across episodes, identify reusable cross-domain structure, store it persistently, and return selected meta-states through a bounded latent bridge. Meta-abstraction and compute leverage are separate claims. Level I requires relational invariance, held-out-domain transfer, causal benefit after reinjection, and reuse across analogous tasks. Level II additionally requires generalization over relations and comparison with a generic support-to-query meta-learner. Compute leverage requires a better held-out utility–compute tradeoff under matched supervision, accessible experience, information/memory budget, and inference compute. Setup and maintenance costs are handled by lifecycle accounting and a break-even horizon. Scaling analyses treat representation-structure measures separately and begin with a no-association null; change-point behavior is secondary. Controls include retrieval, extra test-time compute, latent recurrence, coprocessing, cache processing, reasoning-vector composition, generic meta-learning, coordinate dependence, lexical leakage, and test-memory contamination. No empirical results or claims of consciousness, AGI, superintelligence, or spontaneous intelligence emergence are made.