Cristina Gârbacea

---

Biography

I am Tenure-Track Faculty member and Chief Scientist at CISPA Helmholtz Center for Information Security, where I direct the 💎 CRISTAL Lab — Communication, Reasoning, Intelligence & Safety for Trustworthy ALignment. Our mission is to develop cutting-edge AI systems that can communicate effectively, reason reliably, operate safely in complex environments, and make transparent and trustworthy decisions. For open positions and research opportunities, please see cristal-lab.ai.

Prior to joining CISPA, I was a Postdoctoral Scholar at the University of Chicago, Data Science Institute, working with Prof. Chenhao Tan and the Chicago Human+AI (CHAI) research group. I also collaborated with Prof. Victor Veitch.

I earned my PhD in Computer Science and Engineering from the University of Michigan, advised by Prof. Qiaozhu Mei. My academic background also includes an MSc in Artificial Intelligence (cum laude) from the University of Amsterdam, and a double BSc in Computer Science and Electrical Engineering from Transilvania University of Brasov.

Complementing my academic work, I have completed several Research Scientist internships at Google DeepMind (London) and Microsoft Research (Redmond, Montreal, Cambridge) during my graduate studies.

---

Research Vision & Core Pillars

My research focuses on building human-aligned, safe, trustworthy and continuously adaptable AI systems. As foundation models transition from static text generators to active real-world agents, my work addresses core alignment, safety and adaptability challenges across the following key pillars:

• 🎯 LLM Alignment: Developing principled, scalable methods to ensure foundation models adhere to human intent, values, and context-aware expectations across diverse user preferences.

• 🛡️ AI Safety & Robustness: Developing verifiable alignment mechanisms, safety guarantees, and risk mitigation strategies for foundation models against unexpected failures and adversarial exploits.

• 🤖 Agentic AI & Continual Learning: Investigating autonomous reasoning systems capable of long-horizon planning and tool interaction that adapt over time without safety drift or catastrophic forgetting.

• 📊 Benchmarking & Evaluation: Creating evaluation environments resilient to benchmark contamination to accurately measure evolving AI capabilities and track real-world alignment progress.

---