ELLIS Institute Tübingen PI Sahar Abdelnabi has been awarded funding from Coefficient Giving to support research on developing context-sensitive benchmarks and training methods for sycophancy in AI systems.
Sahar’s research focuses on developing safe, aligned, and steerable AI agents, with an emphasis on security, human aspects, and cooperative multi-agent systems.
Her group, Cooperative Machine Intelligence for People-Aligned Safe Systems (COMPASS), investigates how to build AI systems that are safe, aligned with human values, and robust against adversarial manipulation. Its research spans a broad range of AI safety and security topics, including interpretability, reasoning, evaluation methods, contextual integrity, agentic risks and opportunities, multi-agent dynamics, agents with long-term memory, self-improving agents, alignment, situational awareness, and the study of manipulation and deception.
With this funding, the COMPASS group will focus on developing context-sensitive benchmarks and training methods to better evaluate and address sycophantic behavior in AI systems. The project contributes to ongoing efforts to advance the safety, reliability, and alignment of increasingly capable AI models.
Coefficient Giving is a philanthropic funder and advisor whose mission is to use its resources to help others as much as possible. Coefficient Giving has been particularly invested in AI safety, systematically supporting research in this area.
Find out more about Sahar’s research.
Find out more about Coefficient Giving.