University of Twente Student Theses
How Hard Can It Prompt? Adventures in Cross-model Prompt Transferability
Solovyeva, Olga (2024) How Hard Can It Prompt? Adventures in Cross-model Prompt Transferability.
|
PDF
2MB |
| Abstract: | Recently, the concept of prompting has been introduced as a mechanism to provide con- textual cues to the model, thereby directing its focus towards the downstream task solely through interaction with an input. The concept has evolved into prompt-tuning, a pro- cess in which human intervention is eliminated, and the model autonomously adjusts the prompt parameters in pursuit of optimizing its final performance. Furthermore, the ad- vent of prompt transfer has notably advanced performance in prompt-tuning. However, existing prompt transfer methodologies are primarily based on soft prompts, which lack interpretability and exhibit limited transferability due to their reliance on the model’s ar- chitecture. This study introduces Gradient Guided Discrete Prompt Search (GGDPS), which combines the interpretability and ease of transferability of hard prompts with the expressiveness of soft prompts. GGDPS uses gradients from soft prompts to guide the model in selecting optimal hard tokens, allowing for autonomous prompt generation with- out human intervention. The resulting hard prompts are easily transferable between models and require no architectural modifications. We evaluate the feasibility and applicability of GGDPS, finding that the resulting hard prompts can match or even surpass baseline results on certain datasets. Additionally, we evaluate the transferability of the generated prompts, revealing their success is depended on the pre-trained knowledge of the model. To address this challenge, we introduce Collaborative GGDPS, which enhances generalizability and robustness by creating shared prompts across various models. |
| Item Type: | Essay (Master) |
| Faculty: | EEMCS: Electrical Engineering, Mathematics and Computer Science |
| Subject: | 54 computer science |
| Programme: | Computer Science MSc (60300) |
| Link to this item: | https://purl.utwente.nl/essays/103206 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page