To overcome the limitations of existing adversarial trigger learning (ATLA), this paper proposes Adversarial Trigger Learning with Augmented Objectives (ATLA). ATLA improves the existing negative log-likelihood loss function to a weighted loss function, ensuring that learned adversarial triggers are more optimized for response-type tokens. This allows adversarial triggers to be learned with just a single question-response pair, ensuring good generalization to other similar queries. Furthermore, trigger optimization is enhanced by adding an auxiliary loss function that suppresses evasive responses. Experimental results demonstrate that ATLA outperforms existing state-of-the-art techniques, achieving a near-100% success rate while requiring 80% fewer queries. The learned adversarial triggers also generalize well to new queries and LLMs. The source code is publicly available.
