Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs

TL;DR

This paper introduces ID-SPAM, a novel input-dependent soft prompting method using self-attention, which enhances domain transfer and efficiency in large language models without extensive fine-tuning.

Contribution

It presents a new self-attention based soft prompting technique that dynamically generates prompts based on input tokens, improving transferability and efficiency.

Findings

Outperforms state-of-the-art soft prompting methods

Enhances zero-shot domain transfer capabilities

Maintains low number of trainable parameters

Abstract

The performance of large language models in domain-specific tasks necessitates fine-tuning, which is computationally expensive and technically challenging. This paper focuses on parameter-efficient fine-tuning using soft prompting, a promising approach that adapts pre-trained models to downstream tasks by learning a small set of parameters. We propose a novel Input Dependent Soft Prompting technique with a self-Attention Mechanism (ID-SPAM) that generates soft prompts based on the input tokens and attends different tokens with varying importance. Our method is simple and efficient, keeping the number of trainable parameters small. We show the merits of the proposed approach compared to state-of-the-art techniques on various tasks and show the improved zero shot domain transfer capability.