Reimagining service delivery in the agentic era with Google Public Sector
cloud.google.com·20h ago
AI Research
arXiv CS.AI·May 27·6 min read
Multi-Stakeholder LLM Alignment: Decomposing Estimation from Aggregation
TL;DR
When large language models (LLMs) judge outputs for multiple users with conflicting preferences, current methods blend preference estimation and preference weighting together, creating unstable and unpredictable weight assignments. DecompR separates these two steps using counterfactual calibration to stabilize scoring.
✦ Why It Matters
Engineers can use DecompR to build more stable multi-user systems where conflicting preferences no longer cause unpredictable scoring shifts.
Key Takeaways
- Use DecompR to improve LLM alignment with multiple stakeholders.
- Separate utility estimation from aggregation to reduce instability.
- Implement counterfactual-calibrated weights for better scoring.
- Test outputs with varying stakeholder counts to assess performance.
How It Works
DecompR operates by fixing weights based on the query structure before scoring, which prevents the drift in weights that can occur when candidate scores are aggregated. This separation allows for a more stable estimation of utility for each stakeholder role, leading to more reliable outputs.