New Method Decomposes LLM Alignment for Multi-Stakeholder Tasks

Researchers propose DecompR, a method for aligning large language models (LLMs) in multi-stakeholder scenarios where users have conflicting preferences. Traditional holistic LLM judges conflate utility estimation and aggregation, introducing unstable implicit weights. DecompR fixes counterfactual-calibrated weights from query structure before candidate scoring and estimates per-role utilities independently, removing candidate-dependent weight drift and reducing estimation noise. Experiments show that weighting noise can create large score shifts when stakeholder satisfaction is dispersed, and these shifts increase with stakeholder count. The work is published on arXiv under computer science and artificial intelligence.