Transfer Learning with Evil Twins: Can Obfuscated Prompts Enable Robust Domain Adaptation?

TL;DR: What if evil twins could be used as a form of data augmentation or transfer learning to help models generalize better across domains or tasks?

Research Question: Does training or fine-tuning on evil twin prompts improve the robustness and domain generalization of large language models relative to training on natural prompts alone?

Hypothesis: Exposure to both interpretable and obfuscated (but functionally equivalent) prompts will enhance a model’s ability to abstract task-relevant information, resulting in improved performance on out-of-distribution and adversarial tasks.

Experiment Plan: - Take a downstream NLP or vision-language task (e.g., OOD detection, as in Zhang et al., 2024).

• Construct a dataset of prompt pairs (natural and evil twins).
• Fine-tune or prompt-tune models with a mixture of both.
• Test on standard as well as out-of-domain benchmarks.
• Compare performance, robustness, and sensitivity to adversarial prompts versus models trained on only one prompt type.

References:

• Zhang, Y., Zhu, W.-Q., He, C., & Zhang, L. (2024). LAPT: Label-driven Automated Prompt Tuning for OOD Detection with Vision-Language Models. European Conference on Computer Vision.
• Zhou, K., Yang, J., Loy, C. C., & Liu, Z. (2021). Learning to Prompt for Vision-Language Models. International Journal of Computer Vision.
• Trad, F., & Chehab, A. (2024). Prompt Engineering or Fine-Tuning? A Case Study on Phishing Detection with Large Language Models. Machine Learning and Knowledge Extraction.