Meta’s ads recommendation system is powered by a diverse portfolio of ranking models, each tailored to specific product surfaces and advertiser objectives. These models, though varied, are continuously enhanced through innovations known as “techniques.” Techniques encompass architectural advancements, data representations, training optimizations, and efficiency improvements, but they are highly specific to individual models and do not easily transfer to others.
Due to the complexity and diversity of these models, scaling techniques across the entire fleet requires significant engineering effort. A technique validated on one model rarely applies directly to another, and its shelf life is limited. Thus, the faster proven techniques reach all applicable models, the more value they deliver before being supplanted by new innovations.
To address this challenge, Meta has adopted a multi faceted approach involving standardization and automation. Over the past two years, the company has implemented Standard Model Templates (SMT) to standardize model design and automate the propagation of techniques. This has led to a 92% reduction in per model engineering time and the successful propagation of over 5,000 technique model combinations across the fleet.
However, the process still required significant human bandwidth to make judgment calls, such as determining which techniques to apply to which models and managing the efficient use of compute resources. To overcome these limitations, Meta has introduced AI agents that automate the judgment heavy aspects of technique propagation.
The AI agents operate around three key capabilities: discovery, generalization, and adoption. They leverage a shared knowledge base and cumulative training outcomes to identify applicable techniques, estimate their transferability, and produce prioritized lists of techniques ranked by expected impact. Engineers then review these assessments to make final decisions.
Generalization involves re implementing techniques to fit different models, a process that requires significant customization. The agents start by proposing configurations based on prior training runs, narrowing the hyperparameter search space, and deciding which experiments to continue. This results in a median 4x reduction in training compute. During the adoption stage, agents manage experiments, fine tune models, and recover from failures, significantly reducing the human oversight required.
Agent driven workflows have already shown substantial improvements. Previously, engineers could handle only a handful of models per technique due to the manual effort required. Now, with agents, they can handle three times as many, compressing the time from a proven technique to a shipped model.
Techniques are evaluated through both offline and online metrics. Offline evaluations measure predictive performance and guardrails such as inference costs, while online A/B testing ensures statistical significance before deployment. Agents continuously monitor and prioritize concurrent training jobs to maximize the use of compute resources.
The success of AI agents has led Meta to aim for a significant reduction in the time it takes for techniques to reach all applicable models. The company is currently targeting a reduction from over a year to just three months. This involves further investments in agent intelligence, improving generalization, and optimizing resource allocation.
By agentifying every aspect of the model development lifecycle, Meta envisions a fully automated system that ensures every proven innovation reaches its full potential. Babak Behsaz, Anuj Bindal, and Wenlin Chen are leading this effort, driving Meta toward a future where AI driven innovation can be more efficiently and effectively scaled across its vast model fleet.
Meta is continuing to refine and enhance the AI agents to achieve this goal, focusing on improving end to end throughput and ensuring that the latest techniques are deployed more quickly and efficiently.
Thinking about building an AI product?
Get in Touch