Federated Multi-LLM Inference with Differential Privacy for Cognitive Diagnosis

A heterogeneous federated inference framework enables LLaMA-3.3-70B, GPT-4o-mini, and Claude-3-Haiku to collaborate on diagnostic tasks without accessing raw data. The system employs epsilon-local differential privacy with Laplace noise and residual-based aggregation to manage model heterogeneity.