Directional Preference Alignment with Multi-Objective Rewards Enhances Robustness to Distributional Shifts in User Preferences

This report synthesises findings from 4 peer-reviewed papers addressing the following research question: How does Directional Preference Alignment with multi-objective rewards improve robustness to distributional shifts in user preferences compared to scalar-reward RLHF, as measured by. 11 claims were extracted from source literature; 11 were independently verified against retrieved documents. An automated multi-reviewer quality assessment produced a score of 8.2/10. This report is a machine-generated literature synthesis and does not constitute original research.

Research goal: How does Directional Preference Alignment with multi-objective rewards improve robustness to distributional shifts in user preferences compared to scalar-reward RLHF, as measured by Helpfulness-Harmlessness scores on out-of-distribution evaluations in PaLM 2 variants?

Autonomous literature synthesis. Automated review score: 8.2/10. Full text and citation available at Assignee Research.