What optimization method retains neural network weights but trains a separate encoder for optimal prompts?

The method that retains neural network weights while training a separate encoder for optimal prompts is P-tuning. This technique leverages prompt engineering by adding a trainable prompt, which allows the model to adjust its behavior based on prompt modifications without changing the underlying weights of the pre-trained network. It focuses on optimizing the input specifically for given tasks rather than altering the model itself, making it effective for situations where one wants to produce task-specific outputs while maintaining the integrity of the pre-trained parameters.

In contrast, the other options involve different approaches or techniques that may involve adjusting the model weights or do not specifically focus on the concept of using a separate encoder for prompts in the same way that P-tuning does. Therefore, P-tuning stands out for its unique strategy of prompt optimization while keeping the primary model intact.