Red-teaming large language models typically involves throwing a large library of jailbreak prompts at a target model. But running every attack is expensive, and red-teams often operate under a fixed budget. This paper frames the problem as one of attack scheduling: given a limited number of attempts, which prompts should be tried first?

The proposed approach combines three signals. Static priors encode pre-existing beliefs about which attacks are likely to succeed. Semantic guidance uses the meaning of prompts to avoid redundant attempts or to cover diverse strategies. Online feedback then updates the schedule based on what has already been tried and how the model responded. Together, these steer the red-team toward the most informative attacks within the budget.

The significance is practical: instead of exhaustive testing, red-teams can allocate effort adaptively. The abstract does not report specific results, so the effectiveness of the method remains to be evaluated in full. As a single source, this digest reflects only the paper's own claims.