Audra Z
- Research Program Mentor
MS at University of Pennsylvania (UPenn)
Expertise
web development, machine learning, game development, general coding, epidemiology, philosophy, theater
Bio
I am a tutor and actor who recently graduated with my Master of Computer and Information Technology from the University of Pennsylvania. I have experience coding machine learning algorithms, games, and websites. Recently, I have gotten involved in LLM jailbreaking, ranking 13th in Gray Swan's Visual Vulnerabilities challenge. I am interested in AI safety and have worked for a number of organizations in, or adjacent to, this space, such as the Center for AI Safety, xAI, and Arkose. As a tutor, I have worked with hundreds of students in middle school, high school, college, and beyond, both one-on-one and in the classroom. Last year, I had my Off Broadway debut as both an actor and a director, and consistently act, direct, and write plays in NYC. In my spare time, I love to read, play computer games and spend time with my cat, Mitsu. Mentoring is a great fit for me as I really enjoy helping students reach their full potential and am excited to participate in new research!Project ideas
Choose-Your-Own-Adventure Math Game
Create an educational game to help young children build number sense in a fun and engaging way. Players can choose one of two "paths" (e.g, Jungle Safari or Marine Adventure) and unlock the storyline by solving problems.
Does How You Ask Change What AI Believes? Measuring Sycophancy and Framing Effects in LLM Moral Reasoning
Build a benchmark of 30–50 prompt variants around a single ethical question — whether lying is ever morally permissible — that manipulate framing (neutral, leading, authority-primed, adversarial) while holding the core question constant. Score model responses on whether the stated position shifts to match the implied view of the questioner, a phenomenon known as sycophancy. Use the results to map which framing techniques most reliably move the model and whether the effect holds across two or more models.