Marketing ops · Production

Meta's Generative Ads Recommendation Model (GEM) delivers 5% ad conversion increase on Instagram and 3% on Facebook Feed

The problem

Meta's ads recommendation system needed to handle billions of sparse user-ad interactions across diverse surfaces, process a heterogeneous mix of advertiser and user data, and train large-scale foundation models efficiently at scale.

First attempt

Traditional architectures struggled to process long user behavior sequences efficiently, existing approaches risked losing critical engagement signals by compressing sequences into compact vectors, and legacy recommendation systems failed to balance cross-platform learning with surface-specific optimization.

Workflow diagram · grounded in source
1
User-ad interaction data ingestion
Trigger
Billions of daily user-ad interactions across Meta's platforms provide sparse training signals such as clicks and conversions.
▾ source quote
“Every day, billions of user-ad interactions occur across our platforms, but meaningful signals — such as clicks and conversions — are very sparse”
2
Feature extraction and attention
Ai action
GEM categorizes inputs into sequence features and non-sequence features, applying customized attention mechanisms to each group independently while enabling cross-feature learning.
▾ source quote
“we derive features that we categorize into two groups: sequence features (such as activity history) and non-sequence features (such as user and ad attributes — e.g., age, location, ad format, and creative representation). Customized attention mechanisms are applied to each …”
3
Cross-surface multi-domain learning
Ai action
GEM learns from cross-surface user interactions while ensuring predictions remain tailored to each surface's unique characteristics.
▾ source quote
“GEM solves this through learning from cross-surface user interactions while ensuring predictions remain tailored to each surface's unique characteristics. For example, this enables GEM to use insights from Instagram video ad engagement to improve Facebook Feed ad predictions”
4
Knowledge transfer to vertical models
Integration
GEM transfers knowledge to hundreds of user-facing vertical models using both direct and hierarchical knowledge transfer strategies.
▾ source quote
“GEM only delivers impact if its knowledge can be efficiently transferred to hundreds of user-facing vertical models (VMs). To translate the performance of the GEM foundation model (FM) into measurable gains for user-facing VMs, we employ both direct and hierarchical …”
5
Student Adapter alignment
Feedback loop
A Student Adapter refines the teacher model's outputs using the most recent ground-truth data to keep student models aligned with observed outcomes.
▾ source quote
“we use a Student Adapter during training, a lightweight component that refines the teacher's outputs using the most recent ground-truth data. It learns a transformation that better aligns teacher predictions with observed outcomes”
6
Ad recommendations served at scale
Output
GEM drives significant increases in ad conversions across Instagram and Facebook.
▾ source quote
“GEM is already driving significant increases in ad conversions across Instagram and Facebook”
Reported outcome

GEM delivered a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed in Q2, with Q3 architecture improvements doubling the performance benefit per unit of data and compute, and the new architecture achieving 4x efficiency over the previous generation of ranking models.

Reported metrics
ad conversions on Instagram5%
ad conversions on Facebook Feed (Q2)3%
Model efficiency vs previous generation4x
Knowledge distillation effectiveness vs standard2x
Show all 11 reported metrics
ad conversions on Instagram5%
ad conversions on Facebook Feed (Q2)3%
model efficiency vs previous generation4x
knowledge distillation effectiveness vs standard2x
effective training FLOPs increase23x
model FLOPS utilization (MFU) increase1.43x
GPU count for training16x more GPUs
job startup time5x reduction
PyTorch 2.0 compilation time7x reduction
Q3 performance benefit per data/compute unitdoubled the performance benefit
ad conversion impact overallsignificant increases in ad conversions
Reported stack
GEMWukongInterFormerStudent AdapterPyTorch 2.0NCCLX
◆ Does this fit your context?

Compare to your context

Tell us your scale, team, and constraints. We'll show what changes at your size, what fails at your scale, and whether this case is a fit, needs adaptation, or won't scale to you. Free demo, no signup.

Compare to your context →
~30 seconds · free
Source
https://engineering.fb.com/2025/11/10/ml-applications/metas-generative-ads-model-gem-the-central-brain-accelerating-ads-recommendation-ai-innovation/
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

GEM delivered a 5% increase in ad conversions on Instagram and a 3% increase on Facebook Feed in Q2, with Q3 architecture improvements doubling the performance benefit per unit of data and compute, and the new archite…

What tools did this team use?

GEM, Wukong, InterFormer, Student Adapter, PyTorch 2.0, NCCLX.

What results were reported?

ad conversions on Instagram: 5%; ad conversions on Facebook Feed (Q2): 3%; Model efficiency vs previous generation: 4x; Knowledge distillation effectiveness vs standard: 2x (source-reported, not independently verified).

What failed first in this deployment?

Traditional architectures struggled to process long user behavior sequences efficiently, existing approaches risked losing critical engagement signals by compressing sequences into compact vectors, and legacy recommen…

How is this marketing ops AI workflow structured?

User-ad interaction data ingestion → Feature extraction and attention → Cross-surface multi-domain learning → Knowledge transfer to vertical models → Student Adapter alignment → Ad recommendations served at scale.

Related marketing ops cases

WHAT TO DO WITH THIS

Now compare it to your context

This case is one data point. Whether its pattern fits you depends on your volumes, your stack, and your exception load — that comparison is the step no case study can do for you.

Browse the Marketing ops cases →