The Pulse
Meta says templates cut per-model engineering time by 92%
Meta described a shared-template system for developing and deploying changes across its advertising ranking models. The company says the approach reduced per-model engineering time by 92% over two years.

Meta says a shared-template system reduced the engineering time needed to update each of its advertising ranking models by 92% over two years. In an engineering post, the company says it shipped more than 5,000 technique-model combinations through the approach and improved offline model quality by an average of 0.63% at neutral serving capacity. It also reports a 0.86% cumulative online lift covering 70% of models by count.
The post, published November 18, describes how reusable model designs replace some of the work of adapting each improvement to each model separately. Meta calls the approach Standard Model Templates, or SMT.
Shared designs for ranking models
Meta's ads ranking systems process billions of daily requests across products and optimization goals. As its production models grew in number, they also diverged in architecture, features, training settings and serving constraints.
That diversity made it costly to apply new ideas broadly and complicated reliability and predictions of a change's effect across the fleet, according to the company. An improvement that worked for one model could require further adaptation before it could be used elsewhere.
Templates combine shared components, including architectural choices, feature logic, training settings and serving configurations. Similar models are grouped around a smaller set of designs, while model-specific logic remains where needed.
The templates cover major ranking and retrieval patterns across different model sizes. They also package components for feature selection, feature interaction, model parallelism and hardware-aware serving.
Testing changes across models
Candidate techniques are tested across representative models in a process Meta calls multi-model optimization. The process looks for configurations that improve aggregate performance while limiting regressions, rather than choosing solely on the strength of an average gain.
Candidates are then checked on held-out models to test whether their gains transfer beyond the representative set. Techniques that pass are incorporated into the next template version, which can be rolled out to models using that template.
Meta says it shipped 80 new machine-learning techniques through the workflow over two years. The company also says shared templates make it easier to reuse decisions about resource allocation and cost-performance trade-offs.
Results in the research paper
A related preprint on arXiv describes a study across five iteration cycles spanning 30 months and hundreds of production models. For launched models, the preprint reports a 0.77% average improvement in cross-entropy at neutral serving capacity and a 1.52% cumulative online metric lift.
The authors also report a 92% reduction in per-model iteration engineering time and a 6.3-times increase in technique-model pair adoption throughput. These are the paper's figures for its five-cycle study; the engineering post gives the separate two-year figures cited above.
The paper describes six templates in deployment covering the model fleet, differentiated by model paradigm and feature-interaction type. It compares the template-based process with independent per-model optimization in recommendation systems.
Using agents with templates
Meta also uses the templates as a foundation for a multi-agent approach to model building. At the template level, agents propose optimization ideas, run experiments across representative models, diagnose failures and repeat the process over multiple rounds.
The company identifies further work on template coverage and scaling-law-based optimization, as well as a move toward fully end-to-end agentic workflows. Those workflows include template building, model generation, online evaluation and continuous optimization from online signals.
Sources
Companies in this story
Related stories
- Meta says AI agents handle most ads-ranking technique transfers
- Meta says compute allocation added about 10% to ad ranking on top of model and system advances
- OpenAI names Aleph exclusive ChatGPT Ads reseller in 44 markets
- Meta brings Loblaw shopping to Muse in Canada
- Meta brings Muse AI agent to Australia as local connectors roll out