Source-linked AI summary
Generative Recommendation for Large-Scale Advertising
Ben Xue, Dan Liu, Lixiang Wang, Mingjie Sun, Peng Wang, Pengfei Zhang, Shaoyun Shi, Tianyu Xu, Yunhao Sha, Zhiqiang Liu, Bo Kong, Bo Wang, Hang Yang, Jieting Xue, Junhao Wang, Shengyu Wang, Shuping Hui, Wencai Ye, Xiao Lin, Yongzhi Li, Yuhang Chen, Zhihui Yin, Quan Chen, Shiyang Wen, Wenjin Wu, Han Li, Guorui Zhou, Changcheng Li, Peng Jiang, Kun Gai
TL;DR
Large-scale advertising needs generative recommendation designs beyond standard LLM training and serving recipes. GR4AD co-designs advertisement representation, learning, and inference, and online A/B tests report up to 4.2% higher ad revenue than an existing DLRM-based stack.
Problem
Real-time generative recommendation for advertising requires designs beyond LLM-style recipes because advertising combines complex business information with list-level optimization and online-learning constraints.
Method
GR4AD combines UA-SID, LazyAR, VSL, RSPO, and dynamic beam serving in a production-oriented generative recommender for real-time advertising.
Results
4.2% ad revenue improvement over the existing DLRM-based stack is reported in large-scale online A/B tests, with consistent gains from model and inference-time scaling.
Takeaways & Limitations
GR4AD supports high-throughput real-time advertising recommendation under practical serving budgets and has been deployed to over 400 million users.
Takeaways & Limitations
LazyAR is recommendation-specific and may provide limited speedup for standard LLM serving or long, variable-length generations.
Abstract
from arXiv · showhide
Generative recommendation has recently attracted widespread attention in industry due to its potential for scaling and stronger model capacity. However, deploying real-time generative recommendation in large-scale advertising requires designs beyond large-language-model (LLM)-style training and serving recipes. We present a production-oriented generative recommender co-designed across architecture, learning, and serving, named GR4AD (Generative Recommendation for ADdvertising). As for tokenization, GR4AD proposes UA-SID (Unified Advertisement Semantic ID) to capture complicated business information. Furthermore, GR4AD introduces LazyAR, a lazy autoregressive decoder that relaxes layer-wise dependencies for short, multi-candidate generation, preserving effectiveness while reducing inference cost, which facilitates scaling under fixed serving budgets. To align optimization with business value, GR4AD employs VSL (Value-Aware Supervised Learning) and proposes RSPO (Ranking-Guided Softmax Preference Optimization), a ranking-aware, list-wise reinforcement learning algorithm that optimizes value-based rewards under list-level metrics for continual online updates. For online inference, we further propose dynamic beam serving, which adapts beam width across generation levels and online load to control compute. Large-scale online A/B tests show up to 4.2% ad revenue improvement over an existing DLRM-based stack, with consistent gains from both model scaling and inference-time scaling. GR4AD has been fully deployed in Kuaishou advertising system with over 400 million users and achieves high-throughput real-time serving.
1 Introduction
GR4AD addresses the challenges of deploying generative recommendation in large-scale advertising through recommendation-specific designs across representation, learning, and serving. Online A/B tests report up to 4.2% higher ad revenue than an existing DLRM-based stack, with sub-100ms latency and high-throughput deployment.
- Motivation: Real-time generative recommendation in large-scale advertising requires designs beyond direct reuse of LLM techniques because advertising combines multimodal content, business signals, and list-wise objectives.Advertising creatives combine video, product, and advertiser information, while optimization targets ranked lists and business metrics such as eCPM and NDCG.
- Approach: GR4AD co-designs representation, learning, and serving for production-oriented generative advertising recommendation.Its design includes UA-SID, value-aware online learning, and recommendation-specific serving techniques.
- Approach: UA-SID captures complicated advertisement information, while VSL and RSPO align continual optimization with user interest, business value, and list-level objectives.UA-SID uses fine-tuned multimodal embeddings and multi-granularity quantization; VSL and RSPO address value-aware and ranking-aware optimization.
- Results: 4.2% ad revenue improvement is achieved over the existing DLRM-based stack in online A/B tests.The reported gains are consistent across model scaling and inference-time scaling.
- Results: GR4AD achieves <100ms latency and 500+ QPS per L20 under practical resource budgets and serves over 400 million users in Kuaishou advertising.The system has been fully deployed in the Kuaishou advertising system.
2 Related Works
Prior work established generative recommendation and preference optimization, but advertising systems still lacked architectures and learning strategies tailored to online constraints. GR4AD builds on this literature with a streaming-oriented integration of supervised and ranking-guided preference optimization.
- Generative Recommendation: Generative recommendation encodes items as discrete semantic identifiers and formulates recommendation as next-token prediction.TIGER introduced hierarchical semantic IDs through residual quantization, while later systems improved scalability and integrated retrieval and ranking.
- Advertising Recommendation: Existing generative recommenders largely lack architectures and learning strategies specifically designed for advertising systems under online learning constraints.This gap distinguishes advertising deployment from general generative recommendation progress.
- Preference Optimization: Preference optimization methods such as RLHF and DPO address alignment, but many remain offline and rely on static preference data or fixed rollouts.Subsequent variants target stability or multi-objective learning without eliminating the broader offline-data limitation described here.
- Preference Optimization: Advertising and recommendation systems have adapted reinforcement learning through reward shaping, preference weighting, and hierarchical objectives, but often rely on multi-stage pipelines or naive SFT–RL combinations.These design choices limit adaptability to streaming settings.
- GR4AD Positioning: GR4AD designs an efficient SFT–RSPO integration framework tailored to streaming recommender scenarios.The approach builds on principled joint SFT–RSPO optimization with dynamic weighting or shared objectives.
- Semantic IDs: Semantic IDs address sparsity and cold-start limitations of sequential item IDs by encoding item content into hierarchically quantized identifiers.Prior methods use RQ-VAE-based or clustering-based quantization, including residual and codebook-balancing approaches.
3 Methodology
GR4AD co-designs advertisement representation, generative decoding, and value-aware list optimization for production-scale recommendation. Its methodology combines UA-SID tokenization, LazyAR decoding, and RSPO to address advertising-specific modeling and serving constraints.
- Unified Advertisement Semantic ID: UA-SID represents advertisements with fine-tuned multimodal embeddings, collaborative signals, and multi-granularity business features.The design combines instruction tuning, co-occurrence learning, multi-resolution quantization, and numeric mappings for non-semantic signals.
- Unified Advertisement Semantic ID: Balanced multi-resolution quantization improves codebook utilization, while hash-based mappings reduce collisions caused by business attributes such as account IDs and conversion types.These choices target semantic separability and advertising-specific distinctions that content representations alone may miss.
- Lazy Autoregressive Decoder: LazyAR delays dependence on the previous UA-SID from the first decoder layers to an intermediate layer, then applies the remaining layers autoregressively.The first K layers are computed without the previous-level embedding, which is injected through a lightweight fusion operator before the remaining L−K layers.
- Lazy Autoregressive Decoder: The first K decoder layers can be computed once and shared across beams, reducing sequential work at later UA-SID levels during beam search.Only the remaining L−K layers retain autoregressive dependence, targeting the later levels where beam-search computation is concentrated.
- Efficiency and Scope: LazyAR preserves recommendation quality while doubling inference throughput at a 3L configuration, but its design is less suitable for standard LLM decoding.The reported limitation follows from typical LLM serving using small or no beams and variable-length generations whose later-token difficulty may not decrease.
- Value-Aware Learning: GR4AD uses RSPO to optimize value-based rewards over ranked candidate lists rather than relying only on per-item rewards or logged behavior.RSPO uses candidate eCPM values and directly aligns its objective with ranking NDCG, while its loss is shown to upper-bound NDCGcost.
4 Deployment
GR4AD is deployed as a closed-loop advertising system combining realtime serving, indexing, online learning, and reward feedback. Dynamic beam serving, caching, and beam-search optimizations support high-throughput, low-latency generation under fluctuating traffic.
- Deployment: GR4AD serves over 400 million users with 500+ QPS per L20 and latency below 100ms.Its closed-loop deployment integrates reward estimation, online learning, and realtime indexing for continuous model evolution.
- System architecture: The deployment integrates realtime serving, realtime UA-SID indexing, continual online learning, and value-based reward feedback.These four components form a closed-loop architecture for ranking, item mapping, online updates, and reward estimation.
- Online learning: The reward system generates value estimates for candidate ads and streams them as reinforcement-learning logs for ranking-aware RSPO updates.It supplements online supervised-learning logs because direct realtime exploration can incur performance costs.
- Serving optimization: Dynamic Beam Serving adapts beam computation to improve the efficiency–effectiveness trade-off of multi-step decoding.It combines progressively increasing beam widths across steps with traffic-aware adjustment based on instantaneous load and available compute.
- Serving optimization: Off-peak beam expansion uses available compute for broader hypothesis exploration while preserving peak-time latency and throughput budgets.The active beam scale increases during lower traffic and remains controlled under peak load.
- Serving optimization: Caching reuses recommendations for repeated requests within a bounded interval, reducing inference resource consumption without degrading serving performance.Beam-Shared KV Caching further reduces per-step KV-read complexity from O(B·L) to O(L), while TopK Pre-Cut narrows the search space.
5 Experiments
Experiments show that GR4AD’s value-aware learning, dynamic serving, and LazyAR improve revenue or efficiency, while business indicators and scaling studies demonstrate practical gains. Model size and beam width both exhibit clear scaling effects, and UA-SID optimizations further increase revenue.
- Value-Aware Online Learning: RSPO provides the largest improvement among the value-aware optimization components by aligning generation probabilities with list-wise relative rankings.The unified VSL–RSPO training indicator stabilizes online learning while combining user-interest modeling with value optimization.
- Dynamic Beam Serving: Dynamic Beam Width reduces computation without compromising revenue, while TABS increases off-peak beam width by 60% to improve revenue.The mechanisms jointly balance inference efficiency and revenue under changing traffic conditions.
- Lazy Autoregressive Decoder: LazyAR nearly doubles QPS with only a marginal performance decrease by sharing decoder layers and enabling parallel computation.The reported configuration uses nine decoder layers, with the first six shared across beams.
- Business Indicators: Ad delivery for small and medium-sized advertisers increases by 17.5%, while ad conversion rates improve by 10.17%.Less active users show a 7.28% conversion-rate increase, which the authors associate with improved content-based SIDs and realtime indexing.
- Scaling Laws: GR4AD exhibits scaling laws in both model size and beam width, providing guidance for balancing resource consumption with performance gains.The scaling study evaluates both parameter scaling and inference-time search scaling.
- Scaling Laws: Revenue lift increases monotonically from +2.13% to +4.43% as model size grows from 0.03B to 0.32B parameters at fixed beam width 512.The controlled online A/B tests also observe lower training loss for larger models.
- Scaling Laws: At fixed model size 0.16B, revenue lift rises from +2.33% at beam width 128 to +4.21% at beam width 1024.Production beam selection balances these gains against latency and compute budgets.
- UA-SID Quality: Embedding and quantization optimizations together increase revenue by +0.24%.UA-SID ablations evaluate advertisement-specific embedding quality and quantization properties including compression ratio, collision rate, and codebook utilization.
6 Conclusions
GR4AD is a production-oriented generative recommender for real-time advertising, co-designed across tokenization, architecture, learning, and serving. Online tests report up to 4.2% higher ad revenue with real-time deployment at high throughput.
- GR4AD co-designs UA-SID, LazyAR, VSL, RSPO, and dynamic beam serving for large-scale real-time advertising.These components address advertisement representation, generation efficiency, business-value optimization, and traffic-aware decoding.
- 4.2% ad revenue improvement is reported over a strong DLRM baseline, with gains from both model scaling and inference-time scaling.
- GR4AD achieves high-throughput real-time serving in a fully deployed advertising system.
A.1 Connection Between RSPO and NDCGcost
This appendix connects RSPO's ranking-aware loss to NDCGcost by expressing ranking penalties through reward differences and pairwise ordering indicators. Theorem A.1 states that, ignoring constants, LRSPO optimizes an upper bound of NDCGcost.
- NDCGcost is defined through reward-based gains, positional discounts, and pairwise ordering violations.The formulation uses eCPM-derived rewards, discount terms, and negative samples with lower value than the reference sample.
- Theorem A.1 states that, ignoring constant terms, LRSPO optimizes an upper bound of NDCGcost.
- The proof applies a LambdaLoss inequality and Jensen's inequality to bound the per-sample inner terms.
- The proof introduces a complementary weight and normalizes weights over pairwise samples before completing the bound.
A.2 Efficiency Optimizations
The appendix describes inference optimizations that increase throughput or reduce compute while preserving outputs or limiting quality impact. It also motivates dynamic beam width by allocating wider beams to later layers.
- KV caching improves throughput 2.5×, while Beam-shared KV Caching adds 25% QPS and exceeds 3× total speedup over no caching.
- TopK Pre-Cut selects candidates per beam before global selection, increasing GPU parallelism and reducing comparisons without changing results.
- FP8 inference improves throughput by 50.3% with approximately -0.1% marginal revenue change.
- Result caching serves 27.8% of requests directly within a one-minute time window.
- Dynamic Beam Width is motivated by higher prediction reliability in initial layers and a need for wider beams in later layers.
A.3.1 SID Evaluation Offline Metrics.
The SID evaluation examines semantic representations built from multimodal advertisement information and their ability to distinguish interest groups and product categories. Instruction tuning and co-occurrence learning improve discrimination beyond visual-only or untuned representations.
- SID construction: UA-SID analysis combines product, livestream, virtual-content, and advertisement-video information into semantic representations.The templates emphasize product identity, content themes, persona, engagement, delivered value, and marketing strategy.
- Visualization: The t-SNE visualization marks P* as the clustering center of interest groups.
- Offline evaluation: Instruction tuning and co-occurrence learning enable UAE to distinguish interest groups and finer product categories within an interest group.
- Offline evaluation: QARM cannot link visually different videos referring to the same product because it lacks ad-specific product information.