Sleep–Wake Orchestration in Hierarchical LLM Cohorts

This paper proposes a formal architecture for self-optimizing large language model ensembles that alternate between wake (active inference) and sleep (targeted fine-tuning) phases. Grounded in the Adjoint Projections on Computational Hierarchies framework, the system treats cohorts of models as dynamically rotating through resource-bounded cycles: roughly one-third of compute devoted to retraining on informational gaps—regions of the task space poorly covered by peers yet solvable by higher-level models—while two-thirds remain active in production. The paper defines the underlying mathematics of projection–collapse adjunctions, introduces a gap metric for selecting training data, and presents an implementation plan with canary rollouts, envelope-based resource management, and safety guarantees. The approach connects biological sleep consolidation with continual meta-learning in artificial systems, offering a path toward autonomous, energy-efficient LLM ecosystems.