Modern ads ranking systems update models daily or hourly to track shifting user behavior. However, continual training exposes CVR prediction to heterogeneous label noise from accidental clicks, delayed attribution, and system artifacts. These noisy updates can overwrite long-term user preference patterns, leading to catastrophic
forgetting.

We introduce RoLGaL (Robust Learning with Gauardrail-guided Labeling), a framework for robust continual CVR training. RoLGaL combines label diffusion, guardrail-guided correction, and a multi-head auxiliary architecture to distinguish corrupted supervision from genuine distribution shift. It is implemented as an auxiliary task in multi-task learning, improving representation robustness while preserving calibration and adding zero inference latency.

On the Criteo Attribution dataset, RoLGaL improves AUC by 0.25% and reduces LogLoss by 0.55%. Under simulated label corruption, it achieves 0.31% and 0.34% AUC lifts at 1% and 5% noise, respectively. In large-scale A/B testing at a leading advertising
platform, RoLGaL achieves a statistically significant 0.23% lift in a primary revenue-related business metric with zero inference overhead. It has since been deployed across 10+ production CVR models, demonstrating substantial impact at scale.