GloRank: Generative Reranking for Deep Recommender Systems via Global Identifier Generation

Abstract

As a critical stage in the modern multi-stage recommendation pipeline, list-wise reranking determines the final exposed item sequence and directly shapes the user experience by modeling complex dependencies among candidate items within a list. Existing methods typically formulate this task as selecting indices from the local input list. However, this approach suffers from a semantically inconsistent action space: the same output neuron (logit) corresponds to different items across different samples, preventing the model from forming a stable, item-intrinsic representation. To address this, we propose GloRank (Global Action Space Ranker), a generative framework that shifts reranking from selecting local indices to generating global identifiers. Specifically, we represent items as sequences of discrete tokens and reformulate reranking as a token generation task. This design effectively decouples the scoring mechanism from the order of variable inputs, ensuring that items are evaluated against a consistent global standard. To further strengthen the framework, we adopt a two-stage optimization pipeline: supervised pre-training to initialize the model with high-quality behavioral demonstrations, followed by reinforcement learning-based post-training to directly maximize list-wise utility. Experiments on two public benchmarks and a large-scale industrial dataset, complemented by online A/B testing, demonstrate that GloRank consistently outperforms state-of-the-art baselines and also achieves strong robustness in cold-start scenarios. GloRank has been deployed in production on our large-scale short-video recommendation platform, serving over a hundred million users daily.