HiGR: Efficient Generative Slate Recommendation via Hierarchical Planning and Multi-Objective Preference Alignment

TL;DR

HiGR addresses the scalability and quality gaps in slate recommendation by introducing a hierarchical generative framework that decouples global slate planning from per-item decoding. It combines CRQ-VAE semantic tokenization with prefix-level contrastive learning, a coarse-to-fine Hierarchical Slate Decoder, and an odds-based ORPO-like listwise preference alignment to optimize slate quality directly from implicit feedback. Offline results show HiGR outperforms state-of-the-art by over 10% with a 5x inference speedup, while online A/B tests on a large-scale platform report increases in Average Stay Time, Average Watch Time, and content views, demonstrating strong practical impact. This work advances industrial deployments of generative slate recommendations by delivering both higher-quality slates and efficient inference through global planning and structured tokenization.

Abstract

Slate recommendation, where users are presented with a ranked list of items simultaneously, is widely adopted in online platforms. Recent advances in generative models have shown promise in slate recommendation by modeling sequences of discrete semantic IDs autoregressively. However, existing autoregressive approaches suffer from semantically entangled item tokenization and inefficient sequential decoding that lacks holistic slate planning. To address these limitations, we propose HiGR, an efficient generative slate recommendation framework that integrates hierarchical planning with listwise preference alignment. First, we propose an auto-encoder utilizing residual quantization and contrastive constraints to tokenize items into semantically structured IDs for controllable generation. Second, HiGR decouples generation into a list-level planning stage for global slate intent, followed by an item-level decoding stage for specific item selection. Third, we introduce a listwise preference alignment objective to directly optimize slate quality using implicit user feedback. Experiments on our large-scale commercial media platform demonstrate that HiGR delivers consistent improvements in both offline evaluations and online deployment. Specifically, it outperforms state-of-the-art methods by over 10% in offline recommendation quality with a 5x inference speedup, while further achieving a 1.22% and 1.73% increase in Average Watch Time and Average Video Views in online A/B tests.

Figures (6)

• Figure 1: The overall framework of HiGR. It utilizes CRQ-VAE for semantic tokenization and employs a Hierarchical Slate Decoder to perform coarse-to-fine generation via global planning and specific item decoding. Furthermore, an ORPO-based Preference Alignment module is integrated to iteratively optimize slate quality based on user feedback.
• Figure 2: The pipeline of contrastive semantic ID
• Figure 3: The overall architecture of HSD.
• Figure 4: Performance comparison of different decoding length
• Figure 5: Scaling trends of HiGR on convergence loss and NDCG@5 metric with respect to model size.