Direct Preference Optimization

雷达信号

定义

Direct Preference Optimization (DPO) uses relative preferences between candidate responses to adapt a language model. The paper derives an objective that avoids a separately trained reward model and the reinforcement-learning stage used in some alignment pipelines. Its effectiveness still depends on the quality and coverage of the preference data.

来源

Research paper first submitted to arXiv on 2023-05-29.

语境

Useful when comparing alignment training recipes. DPO changes model behavior through training; it is distinct from simply ranking answers at inference time. Catalog note: this is a source-verified reference entry, not a measured current-trend signal. No search-volume or growth claim is made.

来源

• Direct Preference Optimization: Your Language Model is Secretly a Reward Model