Stefano Ermon

Research

Ermon describes developing principled machine learning methods motivated by concrete real-world applications. With Yang Song, Ermon co-authored early work on score-based generative models, later extended into a framework based on stochastic differential equations.³⁵⁶

Ermon is also a co-author of Direct Preference Optimization, which fine-tunes language models on preference data without a separately trained reward model. Later work applied diffusion to discrete data such as text, and Inception's about page lists papers on masked diffusion and block diffusion among the technologies its team has developed.⁷⁸²

Inception

As CEO, Ermon writes that Inception was started to build fast, efficient AI models and that it launched Mercury, its first diffusion-based LLM, in February 2025. Mercury models generate blocks of text in parallel and refine them over a small number of denoising steps rather than producing one token at a time.¹²

Teaching

At Stanford, Ermon has taught Probabilistic Graphical Models (CS 228), Deep Generative Models (CS 236) and Data for Sustainable Development (CS 325B).³

Work and contributions

• Mercury Proprietary product

  Senior author of the 2025 Mercury report, which presents commercial-scale language models based on diffusion that are trained to predict multiple tokens in parallel, starting with the Mercury Coder models for coding.⁴

• Score-based generative modeling Published research

  Co-author with Yang Song of the 2019 paper on generative modeling by estimating gradients of the data distribution, and of the 2020 paper framing score-based generative modeling through stochastic differential equations.⁵⁶

• Discrete diffusion by estimating data ratios Published research

  Co-author with Aaron Lou and Chenlin Meng of a 2023 paper on discrete diffusion modeling for language by estimating ratios of the data distribution.⁸

Timeline

1. Jul 12, 2019Co-authored the paper on generative modeling by estimating gradients of the data distribution.⁵
2. Nov 26, 2020Co-authored the paper on score-based generative modeling through stochastic differential equations.⁶
3. May 29, 2023Co-authored the Direct Preference Optimization paper on aligning language models with preferences.⁷
4. Oct 25, 2023Co-authored the paper on discrete diffusion modeling by estimating ratios of the data distribution.⁸
5. Feb 2025Inception launched Mercury, its first diffusion-based large language model.¹
6. Jun 17, 2025Was a senior author of the Mercury report on ultra-fast diffusion language models.⁴

Honors and recognition

• ICML 2024 Best Paper Award · International Conference on Machine Learning, 2024³
• ICLR 2022 Outstanding Paper Award · International Conference on Learning Representations, 2022³
• ICLR 2021 Outstanding Paper Award · International Conference on Learning Representations, 2021³
• Sloan Research Fellowship · Alfred P. Sloan Foundation³
• IJCAI Computers and Thought Award · International Joint Conferences on Artificial Intelligence³
• NSF CAREER Award · National Science Foundation³
• ONR Young Investigator Award · Office of Naval Research³

Where they work

• Inception 2 people in the wiki
• Stanford University 40 people in the wiki

Related people

• Aditya GroverAlso at Inception
• Michael CollinsLanguage models · Machine-learning theory
• Diederik P. KingmaImage, video, and 3D generation · Machine-learning theory
• Jascha Sohl-DicksteinImage, video, and 3D generation · Machine-learning theory
• Fabian PedregosaMachine-learning theory · Language models
• Sébastien BubeckLanguage models · Machine-learning theory
• Yang SongImage, video, and 3D generation · Machine-learning theory
• Ju-Chieh ChouLanguage models · Image, video, and 3D generation
• Luke Melas-KyriaziLanguage models · Image, video, and 3D generation
• Beren MillidgeLanguage models · Machine-learning theory