Automated jailbreak attacks typically rely on evaluation feedback to iteratively optimize adversarial prompts. However, the abstract notes that existing methods often lack fine-grained diagnostic information, meaning they may not know exactly why a prompt failed or how to adjust it effectively.

The new approach, called Jev-Guided Strategy Bandit (JSB), is proposed to address this gap. By using multi-dimensional judge feedback, JSB aims to provide more detailed guidance during the optimization process, potentially allowing for more targeted refinements.

The method is described as query-based, which suggests it is designed to minimize the number of interactions needed with the target model. This could make attacks more practical, though the abstract does not provide specific performance numbers or comparisons.

Overall, the work highlights the importance of diagnostic feedback in automated attack generation, but the abstract is limited in scope and does not include experimental details or results.