Core Concepts
Prompt Tuning lacks adversarial robustness but relies on task-specific skill neurons for performance.
Abstract
The study explores the relationship between skill neurons, robustness, and prompt tuning. It highlights the transferability of prompts, identifies task-specific skill neurons, and analyzes model performance under suppression of these neurons. The findings suggest a potential link between model robustness and the activation of specific skill neurons.
Directory:
Introduction
Large language models require parameter-efficient finetuning methods due to computational costs.
Prompt Tuning is a popular method that activates specific neurons for tasks.
Related Work
Parameter-efficient finetuning methods adapt PLMs with few additional parameters.
Prompt Transferability allows prompts to be reused across similar tasks.
Methods
Prompt Tuning involves prepending prompt tokens to model inputs in the embedding space.
Skill Neurons are identified based on predictivity for task labels using tuned prompts.
Experiments
RoBERTa and T5 models are tested on various binary classification tasks after prompt tuning.
Adversarial datasets show decreased accuracy, especially for RoBERTa below chance performance.
Results
Skill Neurons are identified in both models, impacting task performance when suppressed.
Discussion
Adversarial robustness may be related to consistent activation of relevant skill neurons.
Conclusion
Prompt Tuning shows high transferability but lacks adversarial robustness, emphasizing the importance of skill neurons.
Stats
"RoBERTa yields below-chance performance on adversarial data."
"T5 retains above-chance performance in two out of three cases."
"Prompt Tuning leads to high prompt transferability between datasets of the same type."
"Suppressing skill neurons significantly impacts task performance."