Q-Tuning: Continual Queue-based Prompt Tuning for Language Models

TL;DR — This paper introduces a continual prompt tuning method called Q-tuning that, to our knowledge, is the first technique for achieving lifelong learning on extremely long task sequences through prompt tuning.

Abstract

This paper introduces **Q-tuning**, a novel approach for continual prompt tuning that enables the lifelong learning of a pretrained language model on a sequence of tasks. For each new task, Q-tuning trains a task-specific prompt by adding it to the prompt queue consisting of the prompts from older tasks. To better transfer the knowledge of older tasks, we design an ensemble mechanism that reweighs previous prompts in queue with a learnable low-rank matrix that reflects their relevance to the current task. To facilitate training and inference with manageable complexity, once the prompt queue reaches its maximum capacity, we leverage a PCA-based eviction rule to reduce the queue's size, allowing the newly trained prompt to be added while preserving the primary knowledge of older tasks. In order to mitigate the accumulation of information loss caused by the eviction, we additionally propose a globally shared prefix prompt and a memory retention regularization based on the information theory. Extensive experiments demonstrate that our approach outperforms the state-of-the-art methods substantially on both short and long task sequences. Moreover, our approach enables the lifelong learning on an extremely long task sequence while requiring only $\mathcal{O}(1)$ complexity for training and inference, which could not be achieved by existing technologies.