HiGR: Industrial-Scale Hierarchical Generative Slate Recommendation Framework in Tencent

Tencent researchers introduced HiGR, a hierarchical generative framework for slate recommendation that improves both efficiency and quality in large-scale recommendation systems. The system achieves 10% better offline performance and 5x faster inference while delivering measurable gains in user engagement metrics across Tencent platforms.

HiGR represents a significant advancement in industrial recommendation systems by addressing fundamental limitations in applying generative models to large-scale slate ranking. The framework's three-layer approach—structured semantic ID generation, hierarchical decoding, and listwise alignment—solves the practical disconnect between theoretical generative recommendation methods and real-world deployment constraints. Traditional generative approaches struggle with entangled token spaces and inefficient autoregressive decoding over long sequences, but HiGR's prefix-contrastive VAE creates a controllable discrete space that enables coarse-grained planning rather than token-by-token generation.

The technical innovation matters because recommendation systems power engagement across platforms serving hundreds of millions of users. Tencent's 5x inference speedup while maintaining quality improvements addresses a critical bottleneck in production systems where latency directly impacts user experience and platform scalability. The 1.22% watch time increase and 1.73% video plays improvement from online testing demonstrate that efficiency gains don't sacrifice personalization quality.

This work signals a broader industry trend toward hierarchical, multi-objective optimization in recommendation systems. Rather than treating ranking as a single optimization problem, HiGR explicitly balances ranking fidelity, user interest authenticity, and content diversity through ORPO-based alignment. The successful deployment across multiple Tencent surfaces validates the approach at scale, suggesting other major platforms will likely adopt similar hierarchical generative architectures.

Looking ahead, the key question is whether this framework generalizes beyond video recommendation to e-commerce, social feeds, and other domains. Open challenges remain in domain adaptation and handling extreme-scale inventory, but HiGR's modular design suggests promising extensibility.

• →HiGR achieves 10% quality improvement and 5x faster inference through hierarchical generative modeling, addressing scalability limitations in slate recommendation systems.
• →The framework's structured semantic ID space and coarse-grained decoding fundamentally differ from token-level generative approaches, enabling efficient global slate planning.
• →Online A/B testing on Tencent platforms achieved 1.22% watch time increase and 1.73% video plays growth, validating production-scale performance.
• →Multi-objective optimization balancing ranking accuracy, user interest, and diversity through ORPO-based alignment improves recommendation holistic quality.
• →Successful deployment serving hundreds of millions of users demonstrates industrial applicability and likely signals broader adoption of hierarchical generative approaches across the industry.