Reducing Retraining by Recycling Parameter-Efficient Prompts
This work addresses the inefficiency of retraining prompts for parameter-efficient methods in large language models, which is an incremental improvement for users of such models.
The paper tackles the problem of retraining task-specific prompts when a frozen pre-trained large language model is updated, by proposing 'Prompt Recycling' methods that transform prompts from a source model to work with a new target model without costly retraining. The result shows that recycling is possible, with the best settings successfully recycling 88.9% of prompts and outperforming baselines, though performance improvements are still needed.
Parameter-efficient methods are able to use a single frozen pre-trained large language model (LLM) to perform many tasks by learning task-specific soft prompts that modulate model behavior when concatenated to the input text. However, these learned prompts are tightly coupled to a given frozen model -- if the model is updated, corresponding new prompts need to be obtained. In this work, we propose and investigate several approaches to "Prompt Recycling'" where a prompt trained on a source model is transformed to work with the new target model. Our methods do not rely on supervised pairs of prompts, task-specific data, or training updates with the target model, which would be just as costly as re-tuning prompts with the target model from scratch. We show that recycling between models is possible (our best settings are able to successfully recycle $88.9\%$ of prompts, producing a prompt that out-performs baselines), but significant performance headroom remains, requiring improved recycling techniques.