Concentrate Attention: Towards Domain-Generalizable Prompt Optimization for Language Models

TL;DR

This work offers a fresh objective towards domain-generalizable prompts optimization named Concentration, which represents the "lookback" attention from the current decoding token to the prompt tokens, to increase the attention strength on prompts and reduce the fluctuation of attention distribution.

Abstract

Recent advances in prompt optimization have notably enhanced the performance of pre-trained language models (PLMs) on downstream tasks. However, the potential of optimized prompts on domain generalization has been under-explored. To explore the nature of prompt generalization on unknown domains, we conduct pilot experiments and find that (i) Prompts gaining more attention weight from PLMs' deep layers are more generalizable and (ii) Prompts with more stable attention distributions in PLMs' deep layers are more generalizable. Thus, we offer a fresh objective towards domain-generalizable prompts optimization named "Concentration", which represents the "lookback" attention from the current decoding token to the prompt tokens, to increase the attention strength on prompts and reduce the fluctuation of attention distribution. We adapt this new objective to popular soft prompt and hard prompt optimization methods, respectively. Extensive experiments demonstrate that our idea improves comparison prompt optimization methods by 1.42% for soft prompt generalization and 2.16% for hard prompt generalization in accuracy on the multi-source domain generalization setting, while maintaining satisfying in-domain performance. The promising results validate the effectiveness of our proposed prompt optimization objective and provide key insights into domain-generalizable prompts.

Figures (16)

• Figure 1: Domain generalization capabilities across various prompting methods (ICL brown2020language, RL deng2022rlprompt, Soft lester2021power) in sentiment classification tasks.
• Figure 2: Illustration of Concentration. The tokens in the blue square are prompt, and those in yellow are input sequences. Concentration represents the model's attention on prompt tokens in forward pass when decoding <mask> token.
• Figure 3: Left: concentration strength of various prompting methods in the last 5 layers (layers 19 to 23). Right: boxplots of the concentration strength in the last layer. Overall, prompts that exhibit good domain generalization gain higher concentration strength and lower concentration fluctuation. The concentration strength of each layer is shown in Appendix \ref{['apdx:att-distribution']}.
• Figure 4: Framework for Soft Prompt Optimization.
• Figure 5: Framework for Hard Prompt Optimization.