Multi-Stakeholder LLM Alignment: Decomposing Estimation from Aggregation
A new paper on arXiv (2605.26878) addresses the challenge of aligning LLMs with multiple stakeholders who have conflicting preferences. The authors show that holistic LLM judges conflate utility estimation and aggregation, leading to unstable implicit weights. This weighting noise can cause large score shifts, especially when stakeholder satisfaction is dispersed, and these shifts increase with the number of stakeholders. To solve this, they propose DecompR, which decomposes the process: counterfactual-calibrated weights are fixed from query structure before candidate scoring, and per-role utilities are estimated independently. This removes candidate-dependent weight drift and reduces estimation noise. The paper includes empirical and theoretical evidence demonstrating the effectiveness of DecompR in producing more stable and reliable outputs for multi-stakeholder tasks.
Developers building LLM systems with diverse user groups can achieve more stable and fair outputs.