William Overman

PhD Student, Stanford

wpo@stanford.edu

Bio

My research focuses on developing theoretically grounded approaches to AI safety and alignment [NeurIPS'24, NeurIPS'25], with a particular emphasis on AI control and scalable oversight. I'm especially interested in how tools from black-box uncertainty quantification, reinforcement learning, and theoretical computer science can help us reason rigorously about the behavior of powerful AI systems.

I’m a Ph.D. student in Operations, Information, and Technology at Stanford Graduate School of Business, advised by Mohsen Bayati. Before Stanford, I was a visiting researcher at the Institue for Basic Science in South Korea and earned an M.Sc. in Computer Science from UC Irvine. I graduated from Caltech in 2020 with a double major in Mathematics and Computer Science.

In addition to my primary focus on AI safety and alignment, my research also explores reinforecment learning [ICLR'22, RLC'25], causal inference [NeurIPS'24, arXiv'25], and AI applications in healthcare [ SSRN'23,arXiv'25]. Throughout my graduate studies, I have interned at Uber, where I have applied my research work in RL and causal inference to problems in the ridesharing and delivery marketplaces.

Publications

‡ indicates equal contribution. ※ indicates equal contribution, sole student.

The Oversight Game: AI Control in Markov Games

W Overman, M Bayati

Early version in NeurIPS'25 Workshop: ML×OR Workshop. 2025.

Conformal Arbitrage: Risk-Controlled Balancing of Competing Objectives in Language Models

W Overman, M Bayati

NeurIPS'25: Neural Information Processing Systems. 2025.

Paper arXiv

Can We Validate Counterfactual Estimations in the Presence of General Network Interference?

S Shirani, Y Luo, W Overman, R Xiong, M Bayati

arXiv'25: arXiv preprint arXiv:2502.01106. 2025.
Accepted for Oral Presentation at the Conference on Digital Experimentation @ MIT (CODE@MIT), 2025
Accepted for presentation at the MSOM Technology, Innovation, and Entrepreneurship SIG, 2025.

Paper arXiv Code

Aligning Model Properties via Conformal Risk Control

W Overman, JJ Vallon, M Bayati

NeurIPS'24: Neural Information Processing Systems. 2024.

Paper arXiv

Higher-Order Causal Message Passing for Experimentation with Complex Interference

M Bayati‡, Y Luo‡, W Overman‡, S Shirani‡, R Xiong‡

NeurIPS'24: Neural Information Processing Systems. 2024.

Paper arXiv Code

Global convergence of multi-agent policy gradient in markov potential games

S Leonardos‡, W Overman※, I Panageas‡, G Piliouras‡

ICLR'22: International Conference on Learning Representations. 2022.

Paper arXiv Code

Independent natural policy gradient always converges in markov potential games

R Fox‡, SM McAleer‡, W Overman※, I Panageas‡

AISTATS'22: Artificial Intelligence and Statistics. 2022.

Paper arXiv Code

The Oversight Game: AI Control in Markov Games

W Overman, M Bayati

Early version in NeurIPS'25 Workshop: ML×OR Workshop. 2025.

Conformal Arbitrage: Risk-Controlled Balancing of Competing Objectives in Language Models

W Overman, M Bayati

NeurIPS'25: Neural Information Processing Systems. 2025.

Paper ArXiv

Can We Validate Counterfactual Estimations in the Presence of General Network Interference?

S Shirani, Y Luo, W Overman, R Xiong, M Bayati

arXiv'25: arXiv preprint arXiv:2502.01106. 2025.
Accepted for Oral Presentation at the Conference on Digital Experimentation @ MIT (CODE@MIT), 2025
Accepted for presentation at the MSOM Technology, Innovation, and Entrepreneurship SIG, 2025.

Paper ArXiv Code

On aligning prediction models with clinical experiential learning: A prostate cancer case study

JJ Vallon, W Overman, W Xu, N Panjwani, X Ling, S Vij, HP Bagshaw, ...

arXiv'25: arXiv preprint arXiv:2509.04053. 2025.

Paper ArXiv

Aligning Model Properties via Conformal Risk Control

W Overman, JJ Vallon, M Bayati

NeurIPS'24: Neural Information Processing Systems. 2024.

Paper ArXiv

Higher-Order Causal Message Passing for Experimentation with Complex Interference

M Bayati‡, Y Luo‡, W Overman‡, S Shirani‡, R Xiong‡

NeurIPS'24: Neural Information Processing Systems. 2024.

Paper ArXiv Code

Beating price of anarchy and gradient descent without regret in potential games

I Sakos, S Leonardos, SA Stavroulakis, W Overman, I Panageas, G Piliouras

ICLR'24: International Conference on Learning Representations. 2024.

Paper ArXiv

Improved Regret Bound for Safe Reinforcement Learning via Tighter Cost Pessimism and Reward Optimism

K Yu, D Lee, W Overman, D Lee

RLC 2025 (Reinforcement Learning Conference).
Journal version: Reinforcement Learning Journal (2025).

Paper ArXiv

Occupancy Prediction with Patient Data: Evaluating Time-Series, Patient-Level Aggregation, and Deep Set Models

SH Kim‡, W Overman※, J Pauphilet‡, WC Cha

Major Revision at Manufacturing & Service Operations Management (MSOM).

Paper ArXiv

Global convergence of multi-agent policy gradient in markov potential games

S Leonardos‡, W Overman※, I Panageas‡, G Piliouras‡

ICLR'22: International Conference on Learning Representations. 2022.

Paper ArXiv Code

Independent natural policy gradient always converges in markov potential games

R Fox‡, SM McAleer‡, W Overman※, I Panageas‡

AISTATS'22: Artificial Intelligence and Statistics. 2022.

Paper ArXiv Code

Some Ordered Ramsey Numbers of Graphs on Four Vertices

W Overman, JF Alm, K Coffey, C Langhoff

Australasian Journal of Combinatorics, Vol 88(3), 266–281. 2024.

Paper ArXiv

Vitæ

Full Resume in PDF.

Website design from Martin Saveski. Code from this GitHub repo.
William Overman
wpo@stanford.edu