Ranking Augmented Generation: Leveraging Diversity for Better Text Generation

Generative models inherently possess the ability to produce diverse outputs, allowing them to tackle a wide range of tasks and complex scenarios. However, they typically generate only a single solution per task, underutilizing their potential for diversity and often resulting in suboptimal outcomes. To better leverage the inherent diversity of generative models, we propose the Ranking Augmented Generation (RaAG) paradigm. This approach encourages the generation of diverse candidates to promote broader exploration, followed by the selection of the optimal candidate as the final result. Given the widespread availability of strong generative models, the core challenge of our paradigm lies in the effectiveness of the selection process. A common approach involves training a ranker to identify the optimal candidate. However, the primary objective is to select the best candidate, not to produce an accurate ranking of all candidates. This misalignment between the task objective and the training objective hinders overall performance. In this work, we propose a ranker optimized for Top–1 accuracy to enhance the performance of generative models. Specifically, we refine the ranking objective to better align with the task of identifying the best candidate. To enable comprehensive evaluation, our model integrates multiple expert modules, each specializing in a distinct ranking dimension. Our method achieves state-of-the-art Top–1 accuracy on both summarization and translation datasets, outperforming existing rankers. Further analysis confirms that our model effectively augments generative models by leveraging more accurate candidate selection.