Residual correction is essential: removing it while retaining LDP yields MAE worse than the single-LLM baseline across all datasets¶

Claim · All claims
Evidence · 1 study · causal 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? In the ablation study, federation with DP but without residual correction produces MAE worse than the single-LLM baseline on all three datasets, with UCI showing the largest degradation (0.2997 vs. 0.1132). → 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? · causal · r2

Ablation study (Table 4) with six conditions decomposing each component's contribution; the article concludes "naive averaging of heterogeneous LLM outputs without correction is harmful" because heterogeneous LLMs differ in calibration, granularity, and systematic biases.

"removing RC while retaining LDP (ε=2.0) yields MAE worse than the single-LLM baseline across all datasets, with the UCI dataset showing the most dramatic degradation (w/o RC: 0.2997 vs. 0.1132 baseline)"

Discussion¶

Related Claims¶

• Federated aggregation is the primary driver of performance gains, contributing an average 11.99% improvement — related
• Laplace-based local differential privacy outperforms Gaussian noise under Rényi-DP accounting for bounded (0,1)K prediction outputs — related
• Heterogeneous multi-LLM federated inference improves cognitive diagnosis MAE over single-LLM baselines across three educational benchmarks with minimal privacy cost — 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
• Federated improvements over the single-LLM baseline are statistically significant across all three datasets, with Cohen's d from 0.62 to 0.89 — related