Joie Zhang

Hi! I’m Joie Zhang (pronounced as “joey”). Currently, I am a MATS Fellow working with Prof. Stephen Casper on technical AI governance. I am also a recipient of the five-year Future of Life Vitalik Buterin PhD Fellowship.

I recently graduated magna cum laude from Princeton, majoring in computer science, where I was grateful to be co-advised by Prof. Danqi Chen and Prof. Peter Henderson. I was also previously a MATS Fellow working with Dr. Lewis Hammond on multi-agent safety.

My research focuses on alignment + safety, technical governance, and AI law and policy.

1. Alignment + Safety: I’m particularly interested in investigating MARL for safety and safety for MARL: how can we use multi-agent RL to improve the adversarial robustness of language models? Conversely, how can we tackle new safety risks that emerge in multi-agent LLM systems? And, can benign coordination and cooperation training necessarily be decoupled from catastrophic collusion (e.g. agent swarms conducting cyberattacks)?

2. Technical Governance: What should robust and generalizable evals look like for highly-personalized persistent agents like Muse and Instinct? What is going on with data centers, and how can hyperscalers build community relationships with binding commitments? How exactly do we pace the frontier responsibly, cohesively, and continuously?

3. AI Law and Policy: How should tort law adapt in a world with ASI? Does ASI provide a stronger case for evaluating AI agents through the lens of strict liability rather than negligence?

Outside of research, I led AI@Princeton, which is Princeton’s largest undergrad organization focused on increasing AI literacy on campus. I was also the Treasurer of Princeton Women in Computer Science, a CS teaching assistant, and a member of the Edwards Collective. When I’m not queueing jobs or reading outputs, I like to run, swim, and pet strangers’ dogs.

Email: [x]@alumni.princeton.edu where [x]=joie

research highlights:

• Multi-Agent Systems. Scaling Laws for Strategic Interactions (ICML Agents in the Wild Workshop, 2026)
• Agent Safety Evals. Dynamic Risk Assessments for Offensive Cybersecurity Agents (NeurIPS, 2025)
• Reasoning-Focused Long-Context Evals. LongProc (COLM, 2025), Reasoning-Focused Evaluation of Efficient Long-Context Inference Techniques (NeurIPS Efficient Reasoning Workshop, 2025)