LoopCTR: Unlocking the Loop Scaling Power for Click-Through Rate Prediction

A research team led by Jiakai Tang has introduced LoopCTR, a groundbreaking architecture designed to overcome the scaling limitations of Transformer-based models for click-through rate (CTR) prediction. Traditional scaling by adding parameters creates unsustainable computational and storage overhead for industrial deployment. LoopCTR's core innovation is its 'loop scaling' paradigm, which increases training-time computation through the recursive reuse of a compact set of shared model layers. This cleverly decouples the growth in computational power from the growth in actual parameters. The architecture itself is a 'sandwich' design, enhanced with Hyper-Connected Residuals and Mixture-of-Experts (MoE) components, and employs process supervision at every loop depth to encode complex, multi-loop benefits directly into the shared parameters.

This design enables a highly efficient 'train-multi-loop, infer-zero-loop' operational strategy. During inference, the model requires only a single forward pass with zero loops, yet it demonstrably outperforms all existing baselines. Experiments across three public benchmarks and one large-scale industrial dataset confirm its state-of-the-art performance. Furthermore, an oracle analysis reveals an untapped performance headroom of 0.02 to 0.04 AUC, with models trained with fewer loops showing higher potential ceilings. This points toward a promising future for adaptive inference systems that could dynamically adjust their computational depth based on input complexity, optimizing the trade-off between accuracy and latency in real-time ad serving and recommendation engines.