🤖 AI Summary
Prompt engineering remains critical for enhancing the performance and safety of large language models (LLMs) and vision-language models (VLMs), yet fundamental challenges persist—including poor interpretability of prompt effectiveness, limited generalization across tasks, and vulnerability to adversarial attacks.
Method: We propose a hierarchical prompt abstraction framework coupled with task-aware prompt optimization, establishing—for the first time—an interpretable, mechanistic link between prompt efficacy and internal model representations. Our approach synergistically integrates chain-of-thought reasoning, instruction tuning, contrastive prompt learning, and attribution analysis to jointly optimize differentiable prompt embeddings and discrete template search.
Contribution/Results: Evaluated on 12 diverse benchmark tasks, our method achieves an average accuracy gain of 14.3%, attains 92% of full-supervision performance under few-shot and zero-shot settings, substantially reduces data dependency, and significantly improves robustness against adversarial perturbations.
