New method estimates trust in multi-LLM systems using expert judgment

Large Language Model (LLM) ensembles are increasingly used to improve reliability, but existing aggregation methods typically assume all models are equally trustworthy. This assumption breaks down when LLMs vary in capability or are adversarial. In a new paper, Jiawei Zheng and Jiazhen Zhang from [institution not specified] tackle this by formulating multi-LLM aggregation as an uncertainty-aware trust estimation problem. They adapt structured expert judgment from decision theory—specifically Cooke's log weighting—which uses context-aware calibration questions to assess each model's probabilistic prediction quality. The method penalizes overconfident incorrect predictions and rewards well-calibrated experts.

Evaluated on MMLU and MMLU-Pro across homogeneous, heterogeneous, and contaminated expert panels, the approach shows that while all aggregation methods perform similarly in homogeneous settings, Cooke weighting becomes critical under heterogeneity and contamination. It achieves a superior accuracy-reliability balance and remains robust even when unreliable experts are introduced. The findings suggest that multi-LLM aggregation requires not just combining predictions, but calibrating trust under uncertainty. This could make LLM ensembles far more dependable in production, especially when models are of mixed quality or when some may be compromised.