Direct Large Language Model Alignment Through Self-Rewarding Contrastive Prompt Distillation
This addresses the problem of reducing reliance on costly human annotations for AI alignment, though it appears incremental as it builds on existing techniques like DPO and contrastive prompts.
The paper tackles aligning large language models with human expectations without human-annotated data by proposing DLMA, a method that uses contrastive prompts to generate and evaluate preference data with a self-rewarding score, achieving better performance than RLAIF on LLaMA2 models and surpassing RLHF in experiments.
Aligning large language models (LLMs) with human expectations without human-annotated preference data is an important problem. In this paper, we propose a method to evaluate the response preference by using the output probabilities of response pairs under contrastive prompt pairs, which could achieve better performance on LLaMA2-7B and LLaMA2-13B compared to RLAIF. Based on this, we propose an automatic alignment method, Direct Large Model Alignment (DLMA). First, we use contrastive prompt pairs to automatically generate preference data. Then, we continue to evaluate the generated preference data using contrastive prompt pairs and calculate a self-rewarding score. Finally, we use the DPO algorithm to effectively align LLMs by combining this self-rewarding score. In the experimental stage, our DLMA method could surpass the \texttt{RLHF} method without relying on human-annotated preference data.