Eric Mitchell

Also known as Eric Anthony Mitchell

Eric Mitchell is a researcher at OpenAI whose work on language models includes Direct Preference Optimization (DPO), a way to fine-tune a model on human preference data. He studied computer science at Princeton, worked as a research engineer at Samsung's AI Center in New York City, and began a Stanford PhD in 2019 as a Knight-Hennessy Scholar. At Stanford he coauthored the MEND model editor in 2021, the DetectGPT detector for machine-written text in January 2023, and DPO in May 2023, which won an Outstanding Main Track Runner-Up award at NeurIPS 2023. He defended his PhD in June 2024 and joined OpenAI. As of October 2026, he co-leads OpenAI's Post-training Frontiers team with Yann Dubois.

Now at: OpenAI
Badges: Bridge

Timeline

By 2019

Knight-Hennessy Scholar at Stanford

He earns a bachelor's degree in computer science at Princeton and works as a research engineer at Samsung's AI Center in New York City. In 2019, he starts a computer science PhD at Stanford as a Knight-Hennessy Scholar.
October 2021

Fast model editing with MEND

In October 2021, he posts Fast Model Editing at Scale with Charles Lin, Antoine Bosselut, Chelsea Finn and Christopher D. Manning. The paper proposes MEND, a set of small auxiliary networks that use a single input-output pair to make fast, local edits to a pre-trained model's behavior.
2022

Intern at DeepMind London

In the summer of 2022, he spends four months as a research scientist intern at DeepMind in London.
January 2023

The DetectGPT detector

With Yoonho Lee, Alexander Khazatsky, Christopher D. Manning and Chelsea Finn, he posts DetectGPT in January 2023, a zero-shot method that tells LLM-generated text from human writing. It lifts detection of fake news articles from the 20B-parameter GPT-NeoX from 0.81 AUROC for the strongest zero-shot baseline to 0.95, and it appears at ICML 2023.
May 2023

Direct Preference Optimization

He coauthors Direct Preference Optimization in May 2023 with Rafael Rafailov, Archit Sharma, Stefano Ermon, Christopher D. Manning and Chelsea Finn. DPO fine-tunes a language model on human preference data in a single stage of policy training, which removes the need for a separate reward model and for sampling from the model during fine-tuning.
December 2023

A NeurIPS runner-up award

NeurIPS names the DPO paper one of its two Outstanding Main Track Runner-Ups in December 2023 and schedules it as an oral presentation.
2024

A PhD and a move to OpenAI

He defends his Stanford computer science PhD under Chelsea Finn and Christopher D. Manning in June 2024 and joins OpenAI as a researcher. In October 2024, he opens the Hoover Institution's seminar series on AI-generated content detection with a talk on DetectGPT.
As of October 2026

Post-training Frontiers at OpenAI

He co-leads the Post-training Frontiers team at OpenAI with Yann Dubois. The team post-trains and deploys large frontier models such as o1, o3 and GPT-5-Thinking.