LLM-HYPER: Generative CTR Modeling for Cold-Start Ad Personalization via LLM-Based Hypernetworks

LLM-HYPER is a new framework that uses large language models as hypernetworks to generate click-through rate prediction models for cold-start ads without traditional training. The system achieved a 55.9% improvement over baseline methods in offline tests and has been successfully deployed in production on a major U.S. e-commerce platform.

LLM-HYPER addresses a fundamental challenge in online advertising: the cold-start problem where newly launched promotional campaigns lack sufficient user interaction data for effective machine learning models. Rather than relying on traditional training pipelines, the framework leverages LLMs as generative hypernetworks that directly produce CTR estimator parameters using few-shot Chain-of-Thought prompting over multimodal ad content. This approach represents a significant shift in how the advertising technology industry approaches real-time personalization at scale.

The technical foundation combines CLIP embeddings for semantic similarity matching of past campaigns with LLM reasoning capabilities to infer feature-wise model weights. By retrieving contextually similar historical campaigns and formatting them as demonstration examples, the system enables LLMs to reason about customer intent and content relevance without explicit model training. The introduction of normalization and calibration techniques ensures the generated weights produce numerically stable predictions suitable for production environments.

The deployment metrics demonstrate substantial practical value: a 55.9% improvement in NDCG@10 over cold-start baselines in offline evaluation, validated through real-world A/B testing on a top-tier e-commerce platform. Successful production deployment indicates the framework overcomes traditional barriers around model reliability and latency in high-traffic advertising systems. This convergence of generative AI and ad-tech optimization suggests a broader industry trend toward using LLMs for parameter generation in specialized prediction tasks, potentially extending beyond advertising to recommendation systems and dynamic pricing applications.

• →LLM-HYPER uses large language models to generate CTR predictor weights without training data, solving cold-start ad personalization
• →Achieves 55.9% improvement over cold-start baselines and has been deployed in production on major e-commerce platform
• →Framework combines CLIP embeddings for campaign similarity matching with few-shot LLM prompting for parameter inference
• →Normalization and calibration techniques ensure generated weights produce production-ready CTR distributions for numerical stability
• →Represents emerging trend of using LLMs as hypernetworks for generating specialized model parameters in real-time applications