Technical Summary: HiGR - Industrial-Scale Hierarchical Generative Slate Recommendation Framework

1. Problem Statement

Slate recommendation, which presents users with a ranked list of items in a single display, is critical for platforms like video feeds and news aggregators. While recent generative recommendation methods using Semantic IDs (SIDs) show promise, directly applying them to industrial-scale slate recommendation faces three fundamental disconnects:

1. Entangled SID Spaces: Existing semantic ID methods often produce code spaces where prefixes do not reliably capture shared semantics. This "entanglement" confounds high-level list planning and makes it difficult to guide relevance and diversity.
2. Inefficient Autoregressive Decoding: Generating an $M$-item slate requires decoding $M \times D$ tokens (where $D$ is the number of SID tokens per item). Full-sequence autoregressive beam search introduces substantial latency and entangles intra-item semantics with inter-item transitions, obscuring global slate structure.
3. Misaligned Objectives: Standard token-level likelihood training does not directly optimize holistic slate quality. Real-world platforms require optimizing for multiple list-level objectives simultaneously: ranking fidelity, genuine user interest, and intra-list diversity.

2. Methodology: The HiGR Framework

HiGR (Hierarchical Generative Recommendation) addresses these challenges through a co-designed pipeline comprising three tightly coupled components:

A. Prefix-Contrastive Residual Quantized VAE (PCRQ-VAE)

To create a controllable discrete space for planning, HiGR learns structured SIDs using a modified Residual Quantized VAE.

• Global Quantization: Unlike standard RQ-VAE which optimizes layers independently, PCRQ-VAE introduces a global quantization objective. This aligns the aggregated quantized representation with the original latent embedding, preserving informative residuals across codebook layers and preventing residual vanishing.
• Prefix-Level Contrastive Alignment: To inject semantic and collaborative signals, the model applies contrastive learning (InfoNCE loss) to the first $D-1$ codebook layers (the prefixes). This forces semantically or collaboratively similar items to share high-level SID prefixes while reserving the final codebook layer for fine-grained item discrimination.

B. Hierarchical Slate Decoder (HSD)

HSD shifts the generative process from token-level decoding to a coarse-to-fine hierarchical approach to reduce latency and enable global planning.

• Coarse-Grained Slate Planner: Instead of decoding $M \times D$ tokens sequentially, the planner first generates $M$ "preference embeddings." Each embedding represents the intended semantics for a specific position in the slate. This stage models the global structure of the list.
• Fine-Grained Item Generator: A lightweight generator then decodes the specific $D$-token SID sequence for each item, conditioned on its corresponding preference embedding.
• Efficiency: By restricting expensive beam search to short, item-level decoding (conditioned on the pre-planned embeddings) rather than the full slate sequence, HSD significantly reduces inference latency while maintaining global structure.
• Diversity Regularization: The planner includes an intra-list diversity (ILD) term to minimize pairwise similarity among planned preference embeddings, encouraging diverse slates without post-hoc estimation.

C. Listwise Preference Alignment (ORPO)

To bridge the gap between token generation and user experience, HiGR employs a listwise multi-objective alignment mechanism using Odds Ratio Preference Optimization (ORPO).

• Triple-Objective Optimization: The model is optimized against three complementary objectives:
  1. Ranking Fidelity (O1): Ensuring the order matches user engagement.
  2. Genuine User Interest (O2): Filtering out exposure noise by focusing on items with positive feedback.
  3. Slate Diversity (O3): Preventing repetitive recommendations and "information cocoons."
• Preference Pair Construction: Positive slates are constructed from reranked user engagement sequences. Negative slates are generated by permuting the order, replacing items with those receiving negative feedback, or appending semantically similar items to test diversity constraints.
• Reference-Free: ORPO is used as a reference-free method to reduce computational costs and mitigate objective drift, directly aligning the model with key business metrics.

3. Key Contributions

1. HiGR Framework: An end-to-end generative slate recommendation framework that unifies structured tokenization, efficient hierarchical generation, and listwise preference alignment.
2. PCRQ-VAE: A novel tokenization method that enforces prefix-level contrastive alignment, creating a structured SID space where high-level prefixes reliably encode semantics, facilitating controllable planning.
3. Hierarchical Slate Decoder (HSD): A two-stage decoding architecture that decouples global slate planning from item generation, achieving a 5× inference speedup compared to baselines while preserving quality.
4. Triple-Objective ORPO: A listwise alignment mechanism that jointly optimizes ranking, interest, and diversity without requiring an auxiliary reference model.

4. Experimental Results

Offline Performance

• Quality: On both a large-scale industrial Tencent dataset and the public KuaiRec dataset, HiGR outperforms state-of-the-art baselines (including ListCVAE, SASRec, TIGER, HSTU, and OneRec) by over 10% in recommendation quality metrics (NDCG@5, Hit Rate, Recall).
• Ablation Studies:
  • Removing the prefix-level contrastive constraint in PCRQ-VAE increases SID collision rates significantly (from ~2.4% to ~8.7%), confirming the necessity of structured prefixes.
  • Removing the intra-list diversity (ILD) regularization reduces diversity metrics, while removing context embeddings causes a notable performance drop.
  • Jointly optimizing all three objectives (Ranking, Interest, Diversity) yields the best overall performance, demonstrating that accuracy and diversity are complementary.
• Scaling Laws: Experiments across model sizes (0.05B to 2B parameters) reveal a clear power-law relationship between model capacity and performance, suggesting predictable gains through scaling.

Efficiency

• Latency: HiGR achieves a 5× inference speedup compared to OneRec (a strong generative baseline) while maintaining higher accuracy.
• Resource Usage: In online deployment, HiGR reduces GPU demand by approximately 60% compared to OneRec while keeping P99 latency below 50ms.

Online A/B Tests

Deployed on multiple Tencent platform surfaces (news, videos, web novels) serving hundreds of millions of users, HiGR demonstrated significant real-world improvements:

• Watch Time: +1.22%
• Video Plays: +1.73%
• Stay Time: +1.03%
• Request Count: +1.57%

5. Significance and Claims

The paper claims that HiGR successfully bridges the disconnect between item-level generation and holistic slate optimization in industrial settings. By introducing a structured SID space and a hierarchical decoding mechanism, it solves the efficiency and controllability bottlenecks of existing generative recommenders. The framework proves that generative modeling can be scaled to industrial levels, delivering both superior recommendation quality and significant computational efficiency. The successful deployment on Tencent's platforms validates its applicability for serving hundreds of millions of users with complex, multi-objective recommendation needs.