Larger or Smaller Reward Margins to Select Preferences for Alignment?
This work addresses a critical bottleneck in aligning LLMs with human values by improving data selection, though it is incremental as it builds on existing preference learning frameworks.
The paper tackles the problem of contradictory evaluations in preference learning for aligning large language models by introducing the alignment potential metric, which quantifies the gap between implicit and explicit reward margins to estimate alignment potential. Empirical results show that training on data selected by this metric consistently enhances alignment performance, surpassing existing metrics and achieving state-of-the-art results in self-play data generation scenarios.
Preference learning is critical for aligning large language models (LLMs) with human values, with the quality of preference datasets playing a crucial role in this process. While existing metrics primarily assess data quality based on either explicit or implicit reward margins, they often provide contradictory evaluations for the same data. To address this issue, we introduce the alignment potential metric, which quantifies the gap from the model's current implicit reward margin to the target explicit reward margin, thereby estimating the model's potential to align with the preference data. Empirical results demonstrate that training on data selected by this metric consistently enhances alignment performance, surpassing existing metrics across different base models and optimization objectives. Furthermore, our method extends to self-play data generation frameworks, where the metric is used to identify high-quality data within the self-generated content by LLMs. Under this data generation scenario, our method surpasses current state-of-the-art (SOTA) results across various training settings and demonstrates continuous improvements in alignment performance as dataset size and training iterations increase.