Hi! I’m Artemis.

I’m an M.S. student in Computer Science (Artificial Intelligence) at Stanford University, and an engineer in Apple’s Silicon Engineering Group. Previously, I earned my B.S.E. in Electrical and Computer Engineering (Computer Systems) from Princeton University. My work and research focus on quantitative methods for understanding and evaluating large-scale systems.

Artemis Veizi holding her bound Princeton senior thesis, “Towards Machine Learning for Network Optimization: Buffer Sizing via Reinforcement Learning”, beside the Princeton tiger statue.

   

Language Model Benchmark Contamination Leaves a Person-Fit Signature

Under review — AAAI-27

A benchmark score is trustworthy only if the evaluated model has not trained on the test items, yet existing contamination detectors demand privileged access — model weights, token log-probabilities, or the training corpus — that a public leaderboard entry does not expose. We propose a detector requiring only the graded response matrix: which items each model answered correctly. Treating contamination as psychometric item preknowledge, a contaminated model behaves like an aberrant test-taker, succeeding on items too difficult for its true ability because it has memorized them.

Take a look →  ·  Paper (PDF)  ·  Code

   

See my research & projects or read my CV. You can reach me at aveizi (at) stanford (dot) edu, or find me on LinkedIn and GitHub.