Inference-Time Control

A class of techniques that adjust the output distribution of generative models toward a target distribution during the sampling phase—through guidance or correction—without modifying model parameters, thereby avoiding costly retraining.