# Recommender systems

Every feature change in a recommendation system normally needs a full-scale retrain for A/B testing, which is expensive, while a single small-scale run often gives a misleading signal. Wukong, HSTU and LLM-based recommenders are shown to follow the same functional form of scaling law after different features are added, differing only in fitted coefficients. Fitting that law to a handful of small-scale runs and applying statistical inference predicts, with confidence intervals, whether a feature helps, is neutral or regresses at full scale, using a fraction of full-scale compute.

Every feature change in a recommendation system normally needs a full-scale retrain for A/B testing, which is expensive, while a single small-scale run often gives a misleading signal. Wukong, HSTU and LLM-based recommenders are shown to follow the same functional form of scaling law after different features are added, differing only in fitted coefficients. Fitting that law to a handful of small-scale runs and applying statistical inference predicts, with confidence intervals, whether a feature helps, is neutral or regresses at full scale, using a fraction of full-scale compute.

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%.

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%.

A credit-assignment framework that redistributes delayed reward by explanation rather than temporal proximity traces sparse terminal reward back to the actions that actually mattered hundreds of steps earlier, across Crafter, WordCraft, KuaiRand, and an industry-scale generative recommender.

A credit-assignment framework that redistributes delayed reward by explanation rather than temporal proximity traces sparse terminal reward back to the actions that actually mattered hundreds of steps earlier, across Crafter, WordCraft, KuaiRand, and an industry-scale generative recommender.

The work proposes and deploys a self-evolving recommendation system in which Gemini-family LLM agents act as machine learning engineers: an Offline Agent (Fast Loop, waking every 5 minutes and running daily) generates and offline-scores candidate configurations, while an Online Agent (Slow Loop, running daily) ranks candidates and drives A/B experiments; across several YouTube recommendation surfaces the agents autonomously discovered optimizer switches (Adagrad to RMSprop, FTRL), a Gated Path (GLU) architecture, activation refinement, multi-objective reward synthesis, and reward hyperparameter tuning, with multiple online metrics improving at the 95% confidence level and experiment throughput rising from Θ(1)-Θ(10) per week to Θ(100) per week.

The work proposes and deploys a self-evolving recommendation system in which Gemini-family LLM agents act as machine learning engineers: an Offline Agent (Fast Loop, waking every 5 minutes and running daily) generates and offline-scores candidate configurations, while an Online Agent (Slow Loop, running daily) ranks candidates and drives A/B experiments; across several YouTube recommendation surfaces the agents autonomously discovered optimizer switches (Adagrad to RMSprop, FTRL), a Gated Path (GLU) architecture, activation refinement, multi-objective reward synthesis, and reward hyperparameter tuning, with multiple online metrics improving at the 95% confidence level and experiment throughput rising from Θ(1)-Θ(10) per week to Θ(100) per week.