TransX separates recommendation scaling from serving

Recommendation models can draw on long behavior histories, but processing that history for every request makes online computation grow with demand. It remains unclear whether a model can add capacity to improve recommendations without requiring proportionally more online serving work.

TransX separates reusable behavior-history processing from candidate scoring, then tests how each part and its stored memory scale.

The central question is whether improving recommendation quality with more model capacity must proportionally raise online serving cost. TransX makes this testable by separating history encoding, bounded behavior memory, and candidate-conditioned prediction.

In scaling experiments, more encoder computation improved held-out click AUC, which measures how well predictions rank clicked examples above non-clicked ones, with memory resolution and fusion capacity fixed; gains tapered beyond about four layers.

As the one-pass supervision budget increased, relative returns shifted toward fusion, but this also meant more optimizer updates, so the comparison does not isolate supervision quantity from training duration. Larger models did not systematically need finer memory. These patterns support conditional allocation across components in the tested regimes.

These patterns come from the tested TransX architecture, data and supervision regimes, and deployment conditions; they do not establish a universal scaling rule. The production cost accounting is normalized and excludes some absolute costs, so it is not a complete lifecycle-cost analysis.