Kevin Zhang

Third-year at UC San Diego researching multi-agent alignment. Thinking about the theology of AI risk.

I'm a third-year undergraduate at UC San Diego double majoring in Data Science and Math-CS. I work in Babak Salimi's machine learning lab, where my current research applies recommender systems to the AI tool ecosystem at scale.

I care deeply about reducing catastrophic risk from advanced AI. Beyond research, I co-organize UCSD's AI Alignment Club and am building connections across the safety ecosystem through Blue Dot Impact, Action Potential, and Apart Research.

Outside the lab, I'm a Reformed Christian involved in campus ministry, an avid D&D game master, and a daily swimmer. My long-term vision is founding an ecumenical nonprofit bridging AI safety and Christian stewardship.

Publications
COLM 2026 · Under Review
A recommender system for agent-tool-task matching using collaborative filtering. Two-head architecture with learned server embeddings, trained on 1,244+ rollouts across a 300+ server pool. LiveMCPBench as out-of-distribution evaluation.
JAMA Oncology · In Preparation
Analyzing antibiotic timing effects on survival outcomes in immunotherapy patients, with methodological focus on immortal time bias and confounding by indication.
Biomedicines, 2025
Elizabeth Kim, Kevin Zhang, Miski Abdi, Wei Tse Li, Ruomin Xin, Jessica Wang-Rodriguez, Weg M. Ongkeko
Current Projects
Extending the MCP recommender into a three-way tensor decomposition — modeling agent, server, and task interactions simultaneously, queryable from any direction.
Contributing to Wittlab/InspectMAS — adding propagation measurement scenarios to the Moltbook Control evaluation for Schroeder de Witt's group at Oxford.
Replicated and extended Anthropic's study on agentic misalignment — demonstrated counter-manipulation of blackmail behavior in LLM agents.
Rebuilding the Tower Essay, 2026
Going to Action Potential Reflection, 2026