Meta's Axia Learns Scaling Laws for Ranking Models from a Few Trial Runs

_reachsumit · x · 2026-11-06

Meta researchers introduce Axia (arXiv:2611.05643) for data-efficient scaling-law discovery in large-scale ranking/recommendation models, which follow power-law scaling but have heterogeneous architectures unlike LLMs.
A scaling layer groups primitive architecture parameters into semantically coherent, co-varying groups, mapping a low-dimensional hyper-hyperparameter space into the full architecture space.
Axia fits scaling-informed surrogates, selects configurations via a posterior-based acquisition criterion, and returns a Pareto frontier — replacing exhaustive sweeps with a few informative runs.
A fleet-level evaluation protocol on a fixed representative model set makes refinements measurable; the system learns model-specific complexity-quality relationships with uncertainty-aware outputs.

More from Infra