Heterogeneous multi-LLM federated inference improves cognitive diagnosis MAE over single-LLM baselines across three educational benchmarks with minimal privacy cost¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Across ASSIST09, GSM8K, and UCI datasets, the heterogeneous federation achieves an average 11.99% MAE improvement over single-LLM baselines at ε=2.0 with 0.46% average privacy cost. → Yagna Manasa Boyapati 2026

Evidence¶

Yagna Manasa Boyapati 2026¶

Yagna Manasa Boyapati, Chong Yu, Tianyu Jiang, Justin Zhan. (2026). Privacy-Preserving Heterogeneous Multi-LLM Federated Inference for Cognitive Diagnosis. arXiv preprint. https://arxiv.org/abs/2609.02947 doi:10.48550/arxiv.2609.02947

q2 · i? · design · r2

Evaluation across three educational benchmarks (ASSIST09, GSM8K, UCI Student Performance) comparing the federated framework to single-LLM baselines, reporting "11.99% average im- provement with 0.46% privacy cost at ε= 2.0". Per-dataset gains were 14.19% (ASSIST09), 7.39% (GSM8K), and 14.40% (UCI).

"Overall, 11.99% average im- provement with 0.46% privacy cost at ε= 2.0 demonstrates broad generalization across mathe- matics, reasoning, and holistic assessment domains."

Discussion¶

Related Claims¶

• Laplace-based local differential privacy outperforms Gaussian noise under Rényi-DP accounting for bounded (0,1)K prediction outputs — related
• Federated improvements over the single-LLM baseline are statistically significant across all three datasets, with Cohen's d from 0.62 to 0.89 — a narrower finding that bears on this claim
• Residual correction is essential: removing it while retaining LDP yields MAE worse than the single-LLM baseline across all datasets — related
• The framework generalizes under automatically generated Q-matrices without requiring expert annotation — related
• ε=2.0 provides the best privacy–utility balance, while ε=0.5 introduces excessive noise that degrades performance below baseline on ASSIST09 — related
• The framework scales gracefully: MAE degrades by less than 2% from 10K to 100K EdNet interactions with near-linear processing time — related
• Federated aggregation is the primary driver of performance gains, contributing an average 11.99% improvement — possibly the same claim (merge candidate)