Preview abstract
Optimizing large-scale machine learning systems, such as recommendation models for global video platforms, requires navigating a massive hyperparameter search space and, more critically, designing sophisticated optimizers, architectures, and reward functions to capture nuanced user behaviors. Achieving substantial improvements in these areas is a non-trivial task, traditionally relying on extensive manual iterations to test new hypotheses. We propose a self-evolving system that leverages Large Language Models (LLMs), specifically those from Google's Gemini family, to autonomously generate, train, and deploy high-performing, complex model changes within an end-to-end automated workflow. The self-evolving system consists of an Offline Agent (Fast Loop) that performs high-throughput hypothesis generation to optimize for proxy metrics, and an Online Agent (Slow Loop) that validates candidates against delayed north star business metrics in live production. Our agents act as specialized Machine Learning Engineers (MLEs): they exhibit deep reasoning capabilities, discovering novel improvements in optimization algorithms and model architecture, and formulating innovative reward functions that target long-term user engagement. The effectiveness of this approach is demonstrated through several successful production launches at YouTube, confirming that autonomous, LLM-driven evolution can surpass traditional engineering workflows in both development velocity and model performance.View details
Preview abstract
Recommender systems play an important role in YouTube, one of the largest online video platforms across the world. In this paper, we focus on a real-world multitask ranking model for YouTube recommendations.
While most of the recommendation research is dedicated to designing better models to improve user engagement and satisfaction, we found that research on stabilizing the training for such models is severely under-explored.
As the recommendation models become larger and more sophisticated, they are more vulnerable to training instability issues, \emph{i.e.}, the loss diverges (instead of converging) which can make the model unusable, wasting significant resources and blocking model iterations.
In this paper, we share our understanding and best practices we learned for improving the training stability of a multitask ranking model used in production. We show some properties of the model that lead to unstable training and speculate on the cause. Furthermore, we propose an effective solution to improve training stability based on our observations of training dynamics when model training starts to become unstable. Our experiments on a proprietary dataset show the effectiveness of the proposed method over several commonly used baseline methods.View details