Balancing student privacy with accurate cognitive diagnosis remains a key challenge for AI‑driven educational platforms. We introduce a federated inference framework that enables several commercial LLM APIs—such as LLaMA‑3.3‑70B, GPT‑4o‑mini, and Claude‑3‑Haiku—to collaborate without accessing raw student data or proprietary model internals. The system builds on a heterogeneous multi‑LLM architecture; each entity adds Laplace noise to its prediction locally, achieving $\epsilon$‑local differential privacy, before the results are aggregated. Residual‑based aggregation mitigates heterogeneity among models. Under an honest‑but‑curious trust assumption, API providers are presumed not to misuse queries, and the differential privacy layer shields the published diagnostic outcomes from external inference. Rigorous privacy‑utility analysis demonstrates strong privacy guarantees with negligible accuracy loss. Extensive evaluations on three real‑world educational benchmarks confirm the framework’s practicality and cross‑domain generalizability.
Review