Privacy-Preserving Heterogeneous Multi-LLM Federated Inference for Cognitive Diagnosis

Synthesis: Boyapati et al. (2026) propose a federated inference framework in which several commercial large language models cooperate to perform Cognitive Diagnosis without any model seeing raw student data. The system adds ε-local differential privacy noise to each LLM's prediction before aggregation and uses residual-based aggregation to handle model heterogeneity, operating under an honest-but-curious trust paradigm. The authors report strong Privacy guarantees with minimal accuracy loss across three educational benchmarks, arguing the approach is practical and cross-domain generalizable. The paper advances Privacy-preserving, calibrated AI-driven Cognitive Diagnosis for real-world Adaptive Learning settings.

Approach

Traditional Cognitive Diagnosis models centralize sensitive student data on a single server, raising serious Privacy and regulatory concerns (GDPR, FERPA). Boyapati et al. invert this by keeping student data local and shipping only model predictions to a federation of commercial LLM APIs (LLaMA-3.3-70B, GPT-4o-mini, Claude-3-Haiku). Because no provider receives raw inputs or proprietary internals, the framework sidesteps the leak-and-misuse risks of centralized diagnosis.

Key Findings

1. The heterogeneous multi-LLM architecture achieves accurate diagnosis by combining predictions from several models with residual-based aggregation that mitigates heterogeneity across models.
2. Adding Laplace noise locally to each entity's prediction output before aggregation yields ε-local differential privacy that shields published diagnostic results from external inference.
3. The approach assumes an honest-but-curious trust paradigm in which API providers do not abuse submitted queries.
4. Rigorous privacy–utility analysis shows strong privacy guarantees with minimal accuracy loss.
5. Evaluations across three educational benchmarks confirm practical usability and cross-domain generalizability.

Significance for AI in Education

The paper connects the Cognitive Diagnosis research tradition to large language model infrastructure while foregrounding the Privacy constraints that block real-world deployment. It offers institutions a path to privacy-preserving diagnosis that keeps appropriate trust in AI outputs while protecting learners — a dimension under-explored in most CDM work, which assumes centralized access to sensitive student trajectories.

Connected Concepts

• Cognitive Diagnosis
• Privacy
• Trust Calibration
• LLM
• Adaptive Learning
• Generative AI
• Learning Analytics

Connected Articles

• Bayesian Cognitive Diagnosis Personalized Learning Paths — Bayesian Cognitive Diagnosis for Personalized Learning Paths
• LLM Cognitive Diagnosis Handwritten Math — LLM Cognitive Diagnosis of Handwritten Math
• Calibrating Trustworthiness LLM Education 2026 — Calibrating Trustworthiness of LLMs in Education
• Trust Reliance AI Education 2026 — Trust and Reliance in AI Education
• Zhang ML Student Progress Programming 2026 — Machine-Learning Student Progress in Programming

Citation

Boyapati, Y. M., Yu, C., Jiang, T., & Zhan, J. (2026). Privacy-Preserving Heterogeneous Multi-LLM Federated Inference for Cognitive Diagnosis. arXiv:2609.02947.