LLM-Based Ads Recommendation System for Stability and Predictability
A new research paper on arXiv (2605.21969) introduces an evaluation framework for quantifying stability and predictability in ads recommendation systems. Traditional systems optimize for accuracy metrics like recall or NDCG, but with generative AI growth, stability becomes critical. The paper presents an online validated semantic candidate generation framework using fine-tuned Large Language Models (LLMs) to improve robustness against input perturbations, addressing issues like repeatability, cold start, and under-exploration.
Key facts
- Paper arXiv:2605.21969 introduces a new evaluation framework for ads recommendation stability and predictability.
- Traditional systems focus on accuracy metrics like recall and NDCG.
- Generative AI growth increases need for prediction stability.
- Stability defined as robustness to minor input perturbations.
- Lack of stability causes repeatability, cold start, and under-exploration problems.
- Framework uses fine-tuned LLMs for semantic candidate generation.
- Online validation shows significant improvement.
- Published on arXiv as a cross-type announcement.
Entities
Institutions
- arXiv