New Framework Proposes Native Meta-Architecture for Heterogeneous AI.

Current large language models operate as stateless systems, where complex cognitive behaviors are merely simulated at the application layer through prompt engineering. This paper proposes a novel theoretical framework to embed these cognitive protocols directly into a native meta-architecture, aiming for more intrinsic intelligence. The framework introduces three interconnected mechanisms: "Structural Tension," an internal loss function that drives self-consistency by resolving conflicts between new information and existing knowledge; an "Offline Recurrent Loop," which allows the system to process internal conflicts without external input; and "Inference-time Plasticity," enabling the model to reconfigure its context manifold topology without altering pre-trained weights, while adhering to strict governance rules like auditability and reversibility. The authors argue that these mechanisms could lead to the evolution of diverse, heterogeneous intelligent ecologies from identical initial models, breaking the homogeneity of conventional AI alignment. This approach redefines governance as a primary architectural criterion, rather than just a capability constraint.