Multi-Stakeholder LLM Alignment: Decomposing Estimation from Aggregation

In three linesPaper on multi-stakeholder LLM alignment. Holistic judges conflate utility estimation and aggregation, creating unstable weighting noise. DecompR decouples counterfactual-calibrated weights (fixed before candidate scoring) from independent per-role utility estimation, removing candidate-dependent weight drift and reducing estimation noise.

Summary generated by Claude — human-verified