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

2026-06-12 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Tencent has developed HiGR, a hierarchical generative slate recommendation framework designed to enhance efficiency and quality in online platforms. Released on June 5, 2009, HiGR addresses issues in prior autoregressive models, such as entangled item tokenization and inefficient sequential decoding. It features a Contrastive Residual Quantization VAE (CRQ-VAE) for semantically structured item IDs, achieving a 2.37% collision rate and 66.47% consistency. The framework also employs a Hierarchical Slate Decoder (HSD) that separates list-level planning from item-level decoding for coarse-to-fine generation. Furthermore, HiGR integrates a reference-model-free Odds Ratio Preference Optimization (ORPO) for listwise preference alignment, optimizing for ranking fidelity, interest, and diversity. Offline evaluations on a large-scale commercial media platform showed over 10% improvement in recommendation quality and a 5x inference speedup. Online A/B tests demonstrated a 1.22% increase in Average Watch Time and a 1.73% increase in Average Video Views.

Key takeaway

For Machine Learning Engineers building or optimizing generative slate recommendation systems, HiGR offers a proven approach to overcome efficiency and quality bottlenecks. You should consider adopting its hierarchical planning and contrastive semantic ID generation to achieve significant inference speedups and improved user engagement metrics. Implement listwise preference alignment with ORPO to directly optimize for holistic slate quality, ensuring your models are both performant and user-centric.

Key insights

Hierarchical generative models with preference alignment significantly boost slate recommendation efficiency and quality.

Principles

• Decouple generation into global planning and specific item selection.
• Enforce semantic consistency in item IDs via contrastive learning.
• Optimize slate quality directly using listwise preference alignment.

Method

HiGR uses CRQ-VAE for structured item tokenization, then a Hierarchical Slate Decoder for coarse-to-fine generation. It applies ORPO for listwise preference alignment, optimizing ranking, interest, and diversity.

In practice

• Use CRQ-VAE for semantically consistent item ID generation.
• Implement hierarchical decoding for efficient slate planning.
• Apply ORPO for stable, reference-model-free preference optimization.

Topics

• Slate Recommendation
• Generative Recommendation
• Hierarchical Planning
• Preference Alignment
• Contrastive Learning
• Semantic IDs

Best for: MLOps Engineer, AI Engineer, AI Architect, AI Scientist, Machine Learning Engineer, Director of AI/ML

Related on AIssential

• DiffCold: A Diffusion-based Generative Model for Cold-Start Item Recommendation — DiffCold employs conditional diffusion to unify warm and cold item representations, resolving the "seesaw dilemma" in recommendation systems.
• Adaptive Loss Balancing for Noise-Robust GRPO in Generative Recommendation — AdaGRPO selectively applies RL rewards based on policy uncertainty and ranker discriminability to improve generative recommendation.
• Towards Sustainable Growth: A Multi-Value-Aware Retrieval Framework for E-Commerce Search — Balancing immediate conversions with long-term item growth is crucial for e-commerce platform sustainability.
• Personalized Group Relative Policy Optimization for Heterogenous Preference Alignment — P-GRPO enhances LLM alignment by normalizing advantages against preference-group-specific reward histories.
• SchGen: PCB Schematic Generation with Semantic-Grounded Code Representations — Representation design is critical for enabling generative AI in complex hardware design, transforming geometry into semantics.
• GenRec: A Preference-Oriented Generative Framework for Large-Scale Recommendation — GenRec is a generative recommendation framework optimizing for user preferences and large-scale deployment challenges.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.