A new theoretical framework proposes moving cognitive protocols from application-layer simulation into a native meta-architecture for large language models, which are currently stateless and reliant on prompt engineering. The approach introduces three mechanisms: Structural Tension as an endogenous loss function, an Offline Recurrent Loop for self-processing, and Inference-time Plasticity to reconfigure context topology without modifying pre-trained weights.

  • Structural Tension derives from the conflict between new information and existing manifold topology, driving internal self-consistency rather than external reward optimization.
  • The Offline Recurrent Loop enables the system to maintain a dynamic resting potential and digest structural conflicts without external input.
  • Inference-time Plasticity allows reconfiguration of context manifold topology subject to governance invariants like auditability and reversibility.

The authors argue this framework fosters a heterogeneous intelligent ecology by allowing model instances to evolve distinct topological structures through path-dependent tension resolution, while remaining within hard governance rails.