HLLM-Creator: Hierarchical LLM-based Personalized Creative Generation
Proposes HLLM-Creator, a hierarchical LLM framework for efficient personalized content generation, achieving 0.476% CTR increase in industrial ad scenarios.
Key Findings
Methodology
The approach employs a hierarchical architecture comprising Item LLM, User LLM, and Creative LLM. Item LLM encodes product info into embeddings; User LLM aggregates user interaction history into interest vectors; these are clustered via k-means, then matched with ads using a prediction model. During inference, only top-matched clusters generate personalized titles, significantly reducing computation. Chain-of-thought (CoT) reasoning constructs high-quality synthetic training data, ensuring factual consistency. The model's effectiveness is validated in Douyin search ads, with a 0.476% CTR uplift.
Key Results
- In large-scale Douyin ad experiments, HLLM-Creator outperformed baselines, achieving a 0.476% increase in CTR during A/B testing, with inference speed improved by 3-5 times due to clustering and pruning strategies.
- Interest clustering and ad-user relevance prediction reduced computational overhead, enabling deployment at hundreds of millions of users.
- Chain-of-thought data generation pipeline ensured high-quality, factually consistent titles, reducing hallucination issues prevalent in synthetic data.
Significance
This work addresses the scalability and accuracy bottlenecks in personalized content generation, crucial for industrial applications like advertising and recommendation. By integrating hierarchical modeling, interest clustering, and high-quality data synthesis, it enables real-time, large-scale personalized content delivery. The approach advances both theoretical understanding and practical deployment, bridging the gap between research and industry needs, and setting a new standard for scalable personalized AI content systems.
Technical Contribution
The core innovation lies in the hierarchical LLM architecture that efficiently models user interests and content generation, combined with a chain-of-thought based synthetic data pipeline that ensures factual correctness. The interest clustering and ad matching strategies significantly reduce inference complexity, making large-scale deployment feasible. These contributions collectively push the boundary of scalable, accurate personalized content generation, providing a practical solution for industrial-scale AI applications.
Novelty
This is the first integration of hierarchical large language models with interest clustering and matching strategies tailored for industrial-scale personalized creative generation. The use of chain-of-thought reasoning for synthetic data construction to ensure factual accuracy in generated titles is novel. Unlike prior work that relies on simple user attributes or keyword-based interest modeling, this approach captures complex user interests efficiently, enabling real-time deployment at scale.
Limitations
- The model's performance depends heavily on the quality of interest clustering; rapid interest changes or poor clustering may degrade personalization.
- Chain-of-thought data generation still requires manual tuning and may not transfer seamlessly across different domains.
- Cold-start users and new ad scenarios pose challenges for interest modeling and relevance prediction, requiring further adaptation.
Future Work
Future directions include incorporating multi-modal data (images, videos) for richer interest modeling, developing dynamic interest update mechanisms, and extending应用到更多行业如电商、内容推荐。此外，还将优化模型的鲁棒性和迁移能力，以适应不同场景和用户变化。
AI Executive Summary
In the rapidly evolving landscape of digital content, personalized generation remains a key challenge. Existing AI-generated content systems often rely on creator inspiration or basic user attributes, limiting their ability to produce truly tailored content. This gap is especially critical in online advertising, where understanding individual user preferences can significantly enhance engagement and conversion. Traditional methods struggle with scalability, efficiency, and factual accuracy, hindering industrial deployment.
To address these issues, the paper introduces HLLM-Creator, a hierarchical large language model framework designed for large-scale, personalized creative generation. The system comprises three core components: Item LLM encodes product information; User LLM models user interests from historical behaviors; and Creative LLM generates personalized titles. By clustering user interests and employing a user-ad matching predictor, the approach efficiently narrows down relevant user groups, reducing computational costs. The innovative chain-of-thought reasoning pipeline constructs high-quality, factually consistent training data, overcoming the common hallucination problem in synthetic data.
Experimental validation on Douyin search ads demonstrates the model's effectiveness, with a 0.476% increase in CTR during online A/B testing. The system's scalability and efficiency make it suitable for deployment across hundreds of millions of users, marking a significant step forward in industrial personalized content generation. The research not only advances theoretical understanding but also provides a practical, deployable solution for real-world applications. Future work aims to incorporate multi-modal data and dynamic interest updates, broadening the impact of this technology across various digital content domains.
Deep Dive
Abstract
AI-generated content technologies are widely used in content creation. However, current AIGC systems rely heavily on creators' inspiration, rarely generating truly user-personalized content. In real-world applications such as online advertising, a single product may have multiple selling points, with different users focusing on different features. This underscores the significant value of personalized, user-centric creative generation. Effective personalized content generation faces two main challenges: (1) accurately modeling user interests and integrating them into the content generation process while adhering to factual constraints, and (2) ensuring high efficiency and scalability to handle the massive user base in industrial scenarios. Additionally, the scarcity of personalized creative data in practice complicates model training, making data construction another key hurdle. We propose HLLM-Creator, a hierarchical LLM framework for efficient user interest modeling and personalized content generation. During inference, a combination of user clustering and a user-ad-matching-prediction based pruning strategy is employed to significantly enhance generation efficiency and reduce computational overhead, making the approach suitable for large-scale deployment. Moreover, we design a data construction pipeline based on chain-of-thought reasoning, which generates high-quality, user-specific creative titles and ensures factual consistency despite limited personalized data. This pipeline serves as a critical foundation for the effectiveness of our model. Extensive experiments on personalized title generation for Douyin Search Ads show the effectiveness of HLLM-Creator. Online A/B test shows a 0.476% increase on Adss, paving the way for more effective and efficient personalized generation in industrial scenarios. Codes for academic dataset are available at https://github.com/bytedance/HLLM.