Alignment through Meta-Weighted Online Sampling: Bridging the Gap between Data Generation and Preference Optimization

Junming Yang, Ning Xu, Biao Liu +2

Preference optimization is crucial for aligning large language models (LLMs) with human values and intentions. A significant challenge in this process is the distribution mismatch…