When Do Prompting and Prefix-Tuning Work? A Theory of Capabilities and Limitations

Published: 27 Oct 2023, Last Modified: 24 Apr 2024ICBINB 2023EveryoneRevisionsBibTeX
Keywords: prompt, prefix, LLM, fine-tuning, theory
TL;DR: We show that prompting and prefix tuning have limitations that make them less powerful than full fine-tuning and explain how despite these limitations they can work well in practice for some tasks.
Abstract: Context-based fine-tuning methods like prompting, in-context learning, soft prompting (prompt tuning) and prefix-tuning have gained popularity as they often match the performance of full fine-tuning with a fraction of the parameters. Still, there is little theoretical understanding of how these techniques influence the internal computation of the model and their expressiveness limitations. We show that despite the continuous embedding space being more expressive than the discrete token space, soft-prompting and prefix-tuning are strictly less expressive than full fine-tuning. Concretely, context-based fine-tuning cannot change the relative attention pattern over the content and can only bias the outputs of an attention layer in a fixed direction. While this means that context-based fine-tuning techniques can successfully elicit or combine skills already present in the pretrained model, they cannot learn tasks requiring new attention patterns.
Submission Number: 27