How Generative LLMs Are Transforming CPS Advertising Recommendations

1. Introduction

After large language models (LLMs) achieved notable success in natural language processing, the academic community began exploring how generative models can enhance search‑advertising recommendation systems. Existing work can be divided into two categories: (1) using LLMs for data and knowledge augmentation, representation extraction, and prompt‑driven dialogue recommendation without modifying the LLM—these methods cannot directly model massive collaborative signals; (2) modifying LLMs to directly model collaborative signals from large‑scale data via pre‑training or fine‑tuning, which is a frontier direction for scaling search‑advertising and requires more complex engineering. Recent industry progress includes Meta’s GR, ByteDance’s HLLM, Xiaohongshu’s NoteLLM and NoteLLM‑2, and Kuaishou’s OneRec.

2. Business Requirements & Core Technical Points

CPS advertising recommendation primarily targets off‑site users across multiple scenarios. Business requirements include precise perception of user intent, multi‑objective optimization that balances revenue and user activity, and compatibility with diverse scenarios and task data. Three core technologies are emphasized: explicit‑intent perception for controllable product recommendation, multi‑objective optimization of recommendation effectiveness, and the One4All generative recommendation framework, corresponding respectively to instruction‑following fine‑tuning, preference‑alignment, and an end‑to‑end data‑to‑model paradigm.

3. Explicit Intent Perception for Controllable Product Recommendation

The task of landing‑page product recommendation is a crucial form of off‑site advertising, corresponding to the research problem of Trigger‑Induced Recommendation (TIR). Existing solutions fall into three categories:

Implicit modeling of user intent based on historical behavior sequences.

Using trigger items for I2I recall or generating search queries (sku2query) for product retrieval.

Three‑network architectures that separately represent trigger items, model historical behavior, and estimate fusion weights (e.g., DIHN, DIAN, DEI2N, DUIN).

业务需求&现有方案局限性

Solution:

Automatically generate rich intent descriptions from traditional recommendation data, using intent text + historical product semantic ID sequence as input and target product semantic ID as output.

Reconstruct the task paradigm of trigger‑induced recommendation, leveraging generative instruction‑following fine‑tuning to perceive and dynamically control historical behavior and trigger items.

Automated Intent Generation and Evaluation

Input: user behavior data + target product.

Using Few‑shot Prompting and Chain‑of‑Thought (CoT) strategies, the Yanxi‑81B model summarizes, reasons, and predicts the current intent.

Output a triple of summary‑reason‑prediction data.

Self‑Verification is employed to evaluate the generated explicit intent.

Input‑Output Paradigm + Instruction‑Following Fine‑Tuning

Data are organized as "Input: [Prompt]. Output: [Response]" and three additional task types are added on top of sequential recommendation. The input‑output definitions are illustrated below.

Solution Effects

Offline: The controllable intent model improves HitRate and NDCG by 2–3× compared with non‑intent models, while demonstrating strong controllability.

Online: SKUCTR increases by >3%, and SKUCVR, same‑store orders, and same‑store commissions also see significant lifts.

4. Multi‑Objective Optimization of Recommendation Effectiveness

Existing non‑LLM methods include Shared Bottom, MMOE, PLE for balancing multiple tasks, and ESMM for addressing sample selection bias. LLM‑based methods comprise MORLHF and MODPO (RLHF/DPO‑based linear weighting of multiple rewards) and Reward Soups (interpolating weights of multiple LLMs).

Solution

Integrating behavior and price data to improve click‑to‑purchase conversion and overall ad revenue.

基于DPO的偏好对齐算法

Model purchase preference using a click‑to‑purchase prediction model f(click→purchase).

Treat “clicked & purchased” items as positive examples and “clicked & not purchased” items as negative examples, organizing data as "Input: [Prompt]. Output1: [Response+]. Output2: [Response‑]".

Limitations: DPO only considers the relative relationship between positive and negative examples and does not leverage reward values.

基于RiC (Rewards‑in‑Context) 的偏好对齐算法

Offline training incorporates multiple rewards (click, purchase, price, commission) into supervised fine‑tuning, enabling the model to learn strategies under various reward combinations.

Online training enhances data sparsity by generating augmented data near the Pareto front, followed by SFT generation, reward model scoring, and multi‑objective rejection sampling.

During inference, the learned mapping from preferences to rewards allows flexible adaptation to diverse user preferences.

Solution Effects

Offline: HitRate@1 improves by over 10% on multiple datasets.

Online: SKUCTR rises >1.5%, SKUCVR >7%, and same‑store orders and commissions see notable gains.

5. One4All Generative Recommendation Framework

CPS advertising involves diverse business scenarios, requiring a framework that can handle cross‑scenario adaptation and flexible model‑update strategies.

The design emphasizes extensibility, jointly modeling behavior and semantics to improve generalization, while optimizing model‑update policies for real‑time inference.

Solution Effects

Implemented routine online deployment supporting >10 million daily UV for CPS advertising real‑time inference.

Based on One4All, the system now accommodates additional behavior and language‑understanding tasks, enabling integrated search‑recommendation modeling, user‑behavior summarization, and personalized intent inference.

6. Summary and Future Outlook

Interactive recommendation systems (search‑recommendation joint) remain an open frontier; further exploration of generative model capabilities and product‑level redesign are needed.

Multimodal information understanding and generation: leveraging rich images and videos in the upstream pipeline can enhance recommendation quality and presentation richness.

7. References

Xu L, Zhang J, Li B, et al. Prompting large language models for recommender systems: A comprehensive framework and empirical analysis. arXiv preprint arXiv:2401.04997, 2024.

知乎《一文梳理工业界大模型推荐实战经验》. 2024.

Zhai J, Liao L, Liu X, et al. Actions speak louder than words: trillion‑parameter sequential transducers for generative recommendations. Proceedings of the 41st International Conference on Machine Learning, 2024.

Chen J, Chi L, Peng B, et al. HLLM: Enhancing sequential recommendations via hierarchical large language models for item and user modeling. arXiv preprint arXiv:2409.12740, 2024.

Zhang C, Wu S, Zhang H, et al. NoteLLM: A retrievable large language model for note recommendation. Companion Proceedings of the ACM Web Conference 2024, 2024.

Zhang C, Zhang H, Wu S, et al. NoteLLM‑2: Multimodal large representation models for recommendation. arXiv preprint arXiv:2405.16789, 2024.

Deng J, Wang S, Cai K, et al. OneRec: Unifying Retrieve and Rank with Generative Recommender and Iterative Preference Alignment. arXiv preprint arXiv:2502.18965, 2025.

Ma J, Xiao Z, Yang L, et al. Modeling User Intent Beyond Trigger: Incorporating Uncertainty for Trigger‑Induced Recommendation. Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, 2024.

Shen Q, Wen H, Tao W, et al. Deep interest highlight network for click‑through rate prediction in trigger‑induced recommendation. ACM Web Conference 2022, 2022.

Xia Y, Cao Y, Hu S, et al. Deep intention‑aware network for click‑through rate prediction. ACM Web Conference 2023, 2023.

Xiao Z, Yang L, Zhang T, et al. Deep evolutional instant interest network for CTR prediction in trigger‑induced recommendation. ACM International Conference on Web Search and Data Mining, 2024.

Ma J, Zhao Z, Yi X, et al. Modeling task relationships in multi‑task learning with multi‑gate mixture‑of‑experts. KDD 2018.

Tang H, Liu J, Zhao M, et al. Progressive layered extraction (PLE): A novel multi‑task learning model for personalized recommendations. ACM RecSys 2020.

Ma X, Zhao L, Huang G, et al. Entire space multi‑task model: An effective approach for estimating post‑click conversion rate. SIGIR 2018.

Zhou Z, Liu J, Shao J, et al. Beyond One‑Preference‑Fits‑All Alignment: Multi‑Objective Direct Preference Optimization. ACL 2024.

Li K, Zhang T, Wang R. Deep reinforcement learning for multi‑objective optimization. IEEE Transactions on Cybernetics, 2020.

Rame A, Couairon G, Dancette C, et al. Rewarded soups: towards Pareto‑optimal alignment by interpolating weights fine‑tuned on diverse rewards. NeurIPS 2023.

Rafailov R, Sharma A, Mitchell E, et al. Direct preference optimization: Your language model is secretly a reward model. NeurIPS 2023.

Wu J, Xie Y, Yang Z, et al. beta‑DPO: Direct Preference Optimization with Dynamic beta. NeurIPS 2025.

Lin X, Chen H, Pei C, et al. A Pareto‑efficient algorithm for multiple objective optimization in e‑commerce recommendation. ACM RecSys 2019.

Hu J, Tao L, Yang J, et al. Aligning language models with offline learning from human feedback. arXiv preprint arXiv:2308.12050, 2023.

Yang R, Pan X, Luo F, et al. Rewards‑in‑Context: Multi‑objective alignment of foundation models with dynamic preference adjustment. ICML 2024.