Ari Holtzman

Ari Holtzman is an assistant professor at the University of Chicago whose work spans text generation, language-model behavior, and AI alignment. He studied computer science and philosophy of language at NYU, completed a computer-science PhD at Washington, and helped build the UW team’s 2017 Alexa Prize-winning Sounding Board. He and his collaborators introduced Nucleus Sampling, coauthored QLoRA, and developed MUSE to evaluate language-model unlearning. His 2023 dissertation on interpretation errors in generative models received the William Chan Memorial Dissertation Award. As of October 2026, he is an assistant professor at UChicago; he taught a graduate course on transformers and mechanistic interpretability in Winter 2026.

Now at: University of Chicago
Works: 72
Citations: 8,854
h-index: 23
Badges: Rising

Publication numbers from OpenAlex, as of 18 October 2026.

Timeline

2015

An NYU degree and honors

He completes an interdisciplinary degree at NYU in computer science and philosophy of language. In 2015, he receives Phi Beta Kappa recognition and is named an NYU University honors scholar.
2017

Sounding Board wins the Alexa Prize

As an Allen School doctoral student at the University of Washington, he works on Sounding Board, a conversational agent that wins Amazon’s inaugural Alexa Prize in 2017.
2019

Nucleus sampling

In 2019, he and his coauthors introduce Nucleus Sampling, which truncates the unreliable tail of a language model’s probability distribution to produce more diverse text.
2023

QLoRA and a doctorate

In 2023, he coauthors QLoRA, which reduces memory use enough to fine-tune a 65-billion-parameter model on one 48GB GPU while preserving full 16-bit task performance. He earns a computer-science PhD at the University of Washington and receives the William Chan Memorial Dissertation Award for his work on interpretation errors in generative models.
July 2024

Joining the University of Chicago

In July 2024, he joins the University of Chicago as an assistant professor of computer science and data science. His work focuses on generative models of text and how people can understand them.
2024

A machine-unlearning benchmark

In 2024, he coauthors MUSE, a benchmark that evaluates language-model unlearning across six goals, including preventing privacy leakage, preserving utility on unrelated data, and supporting repeated removal requests.
2025

Alignment and model reliability

In 2025, he studies how to align AI systems with human values and ensure reliable, safe behavior. He probes language models through creative prompting, targeted fine-tuning, model surgery, and analysis of their internal mechanisms.
As of October 2026

Assistant professor at UChicago

As of October 2026, he is an assistant professor of computer science and data science at UChicago. In Winter 2026, he teaches a graduate course on transformers and mechanistic interpretability.