Conventional federated learners assume a global mean, creating a "one-size-fits-all" regime that suppresses minorities under cross‑site heterogeneity. To close this gap, we introduce Fed-Universe, a generalizable S‑G‑T‑H architecture that turns passive averaging into an active negotiation for decentralized alignment.
The S‑layer deploys an edge‑capable large language model (LLM) as a semantic surrogate, translating heterogeneous inputs into standardized patient‑profile prompts.
The G‑layer enforces geometric quality assurance (GQA) via a mapping $f:\Theta \rightarrow Q$ from the statistical parameter space to an interpretable quality domain, using a cosine‑similarity gate to retain minority updates based on geometric rather than statistical proximity.
The T‑layer performs topological Pareto control, identifying a sensitivity‑based knee point $\lambda_k$ to balance multi‑objective aggregation without degrading the majority.
The H‑layer operationalizes a cognitive twin dynamic, mirroring the user's real‑time psychological state and dynamically adapting human‑computer interaction modes through active intent alignment.
To ground the S‑G‑T‑H core empirically, we built a binational synthetic clinical network simulation. The proof‑of‑concept demonstrated a "zero‑sum escape": a minority node can achieve independent gains while preserving overall performance.
Review