Meta Enhances Ad Retrieval Efficiency with New System

What happened

Meta's Andromeda system has achieved near ranking-level performance in ad retrieval, processing over one billion ads per second. This marks a significant improvement in the system's ability to evaluate and score ads efficiently at scale.

The new system's throughput for ads retrieval-stage models is 100 times higher than the 2024 version, indicating a major leap in processing capabilities. This advancement allows the system to handle a much larger volume of ads with greater speed and accuracy.

Why it matters

The increased efficiency of ad retrieval is expected to enhance user experience by delivering more relevant advertisements. This could lead to higher engagement and better business outcomes for Meta's platforms.

Key details

Since 2025, Meta Andromeda has contributed to a 2.4% increase in conversions on Facebook, demonstrating the practical impact of the system's improvements. The expansion of model capacity by four orders of magnitude in the 2024 deployment highlights the system's growth in handling complex tasks.

Generative AI tools have played a crucial role in expanding Meta's ad candidate pool to hundreds of millions of new ads. The new Retrieval Mixture-of-Experts architecture achieves ranking-level performance with 300 times fewer inference-time FLOPs than the 2024 model.