William Overman ✈️ ICML 🇰🇷

@stanford PhD AI safety.

[William Overman ✈️ ICML 🇰🇷 @willoverman]

Generally happy to talk anything AI safety, multi-agent RL/systems, AI governance, or whatever is on your mind. DMs open.

and I used to live in Korea for a while and still come often, so happy to share any recs from cool places to see to foreigner-friendly skincare clinics

[William Overman ✈️ ICML 🇰🇷 @willoverman]

Arriving in Seoul for ICML 2026 🇰🇷

Two posters Thursday at the main conference, both on AI safety & scalable oversight, plus a workshop poster Friday at @taig_icml

[William Overman ✈️ ICML 🇰🇷 @willoverman]

As we race towards a world with mixed human + agent populations, if policy makers intervene on agents (or humans), how can they accurately measure this effect given this additional issue of a mixed population? In https://t.co/GjYN7bnsxI we introduce an estimator for this

[William Overman ✈️ ICML 🇰🇷 @willoverman]

And I’ll give a spotlight at the ML×OR workshop https://t.co/ykU3wbWUfC Saturday morning on our new work The Oversight Game: Learning to Cooperatively Balance an AI Agent's Safety and Autonomy https://t.co/Paucs6sCpg

[William Overman ✈️ ICML 🇰🇷 @willoverman]

I’ll be at #NeurIPS2025 in San Diego from Wed-Sat

If you want to chat AI safety, alignment, or causal inference, feel free to DM!

[William Overman ✈️ ICML 🇰🇷 @willoverman]

Friday morning poster https://t.co/yChVmhwmId on Conformal Arbitrage (a conformal-calibrated post-hoc alignment layer for enforcing safety)

https://t.co/YxDgn0bFoq

[William Overman ✈️ ICML 🇰🇷 @willoverman]

this framework is designed to tackle potential safety risks introduced post-training

it provides a model of control in which the agent learns to defer when risky and otherwise act autonomously as much as possible, while keeping the pretrained policy untouched

[William Overman ✈️ ICML 🇰🇷 @willoverman]

first time sharing a paper on X, but I'm particularly excited about this one

we model AI control as a Markov Game between a Human and a "Superintelligence" (SI)

at each step, the SI chooses "play" or "ask" while the Human chooses "trust" or "oversee"

https://t.co/Paucs6sCpg

[William Overman ✈️ ICML 🇰🇷 @willoverman]

we show that when this game qualifies as a Markov Potential Game (+an additional assumption), we can obtain a certain notion of "alignment"

any decision by the agent to act more autonomously ("play") that benefits itself is formally guaranteed to not harm the human's value