Meta hands ranking-model scaling laws to an automated search: 34 shipped iterations average 0.66% lower normalized entropy and 0.36% cumulative online lift
Synopsis
Meta's ads ranking and retrieval fleet no longer allocates capacity by hand, one model at a time. Previously each scaling decision cost days to weeks of training, and the architecture dimensions are coupled, so they cannot be enlarged independently. Axia maps embedding dimension, interaction depth and interaction width through a scaling layer into complete architectures, then uses Bayesian surrogates to pick a small number of training trials. Across 34 model iterations in two development cycles, offline normalized entropy improved by a mean 0.66% relative and the online key metric by a cumulative 0.36%.
Interpretation
Capacity allocation for ranking models moves from per-model manual trials to a reusable automated workflow that learns each model's own complexity-quality relationship on Meta's ads ranking and retrieval fleet. That allocation was previously made by hand, one model at a time, at a cost of days to weeks of training per decision, while ranking capacity is spread across several coupled architectural dimensions and cannot be summarized by parameter count alone. Across 34 model iterations in two development cycles, each selected configuration replaced its deployment baseline and delivered a mean 0.66% relative offline normalized-entropy improvement and a cumulative 0.36% online key-metric lift; the two cycles gave 0.66% over 11 models and 0.67% over 23 models.
The workflow compresses architecture parameters into a low-dimensional control space and uses uncertainty-aware batch selection to decide which configuration to train next. The scaling layer groups primitive architecture parameters by semantics, and each group is driven by a single control, so a model can be instantiated coherently at any target scale without becoming mis-proportioned. The production parameterization uses three coordinates, interaction depth, interaction width and embedding dimension; under a matched budget of 20 full training runs (16 initialization plus four adaptive), the three-dimensional parameterization attained the strongest frontier on the representative model, while six dimensions scored lower.
Candidate generation replaces exhaustive sweeps with scaling-informed surrogates and switches acquisition criterion with the deployment goal. Comparisons include random Sobol sampling, default Bayesian optimization, Bayesian optimization with cost profiling, an open-loop surrogate and two grid searches, all under matched budgets. Under a 160 MFLOPs cap, Axia's best feasible normalized entropy is 0.070%, 0.059% and 0.042% lower than random, default Bayesian optimization and the open-loop surrogate, with hypervolume 24.5%, 19.0% and 7.9% higher; against Bayesian optimization with cost profiling, normalized entropy is 0.041% lower and hypervolume 1.0% higher.
Perspective
The demonstrated scope is two completed development cycles covering 34 model iterations on Meta's ads ranking and retrieval fleet; a third cycle is in progress and not reported. The controlled ablations use a single representative model, a matched budget of 20 full training runs, and a cost ceiling of 160 MFLOPs per example.
The 0.36% online figure sums the two cycles' shipped lifts over disjoint model sets rather than a separately measured aggregate effect; offline outcomes have a standard deviation of 0.59 percentage points, and one implementation defect in the Bayesian-optimization-with-cost-profiling arm may, as the authors note, overstate the contribution of Axia's remaining changes.