The correct answer is B because model drift is an operational issue where production model performance changes as data, user behavior, or business conditions change. NVIDIA’s recommender systems best-practices documentation states that production modules should be continuously monitored: “Modules are continuously monitored so that the quality of the recommendation can be measured in real time through a range of KPIs.” It further explains that these modules “trigger full retraining should model drift occur, such as when certain KPIs fall below known established baselines.”

NVIDIA’s TAO Toolkit guidance also supports retraining as the correct response to drift: “To avoid model drift or to accommodate changing business requirements, retrain your model regularly.”

Why the other options are incorrect: Increasing hardware resources may improve throughput or latency, but it does not fix degraded model accuracy caused by drift. Permitting input distributions to change without controls is a cause of drift, not a mitigation. Assuming a model will generalize to any data is not a valid AI operations practice. The verified best practice is to monitor deployed models and retrain or update them with fresh, representative data.