Fine-Tuned LLM as a Complementary Predictor Improving Ads System

Researchers demonstrate a novel approach to advertising systems by using fine-tuned large language models as complementary predictors for advertiser forecasting rather than traditional ranking roles. Deployed in production-scale environments, this method improves candidate generation and downstream ranking by leveraging LLM knowledge to predict likely advertisers from user data, delivering measurable offline and online business improvements.

The integration of large language models into recommendation systems represents an ongoing frontier in AI applications, but practical deployment in advertising at scale remains challenging. This research addresses a specific gap by repositioning LLMs from direct rankers or retrieval mechanisms into auxiliary predictive roles. Rather than having LLMs compete with optimized ranking algorithms, the approach uses fine-tuned open-source models to forecast advertiser likelihood based on user profiles and history, feeding this signal upstream to enhance both candidate generation and final ranking stages.

The architectural choice reflects pragmatic engineering constraints in production environments. Advertising systems require extreme latency efficiency and reliability guarantees that generative approaches struggle to meet. By constraining LLM usage to advertiser prediction—a narrower, more bounded task—the system captures language model benefits while maintaining performance requirements. The evidence of measurable online business impact validates the hypothesis that targeted auxiliary predictions unlock system-wide improvements beyond single-stage optimization.

This work suggests a broader pattern emerging in AI: specialized models excel when assigned complementary rather than primary roles. For the recommendation systems and advertising technology sectors, this indicates LLMs can drive value without replacing existing infrastructure. The approach lowers barriers to adoption since it integrates with conventional systems rather than demanding architectural overhauls. The scalability demonstration matters particularly for platforms managing billions of daily predictions, where theoretical improvements must translate to practical feasibility and business metrics.

• →Fine-tuned LLMs function effectively as auxiliary predictors in advertising systems rather than primary rankers or retrievers.
• →Complementary LLM predictions improve both upstream candidate generation and downstream ranking stages simultaneously.
• →Production-scale deployment demonstrates measurable online business impact beyond offline metrics.
• →Constrained LLM roles optimize for latency and reliability requirements of real-world advertising systems.
• →This architectural pattern suggests specialized auxiliary predictions unlock system-wide gains in recommendation applications.