PermameTeiming: Exploring the introduction of persons can improve the effectiveness of AI red-combining

The paper was welcomed at workshop on Regatable ML (Reml) in Neurip 2025.

The latest developments in the management and safety system has called on red partnerships that may ensure effective risks caused by AI models. Many of these telephones emphasize that the ownership and backgrounds can create their role-participation strategies, and therefore the types of risks are likely to produce. While red integration methods promise to assist the red integration of human reductions by enabling high-quality assessment tests, current methods do not view the role of ownership. As a first step to include the Automated Red-All-All ownership, we develop and evaluate the novel method, PersonIAIMING, introduced Personas to the fastest generation process for a powerful veil. In particular, we first notify the modification of the “red-act-acting expert” or “normal AI” person. We then develop the productive algorithm of automatically producing Persona types to suit different seed services. In addition, we develop a collection of new metrics to clearly modify the “Distance of Changes” to achieve existing Affersarial Prompts. Our exams reflect promising development (up to 144.1%) in the achievement price of Personality, while keeping fast distinctions, compared to red red rain. We discuss the power and restrictions of various forms of Persona types and methods of conversion, it enhances the opportunities for the future assessment between the default and the red paths of a person.

• 40 Carnegie Mellon University
• 5
• ** Work done while in Apple