Stefano Ermon researches generative models and machine learning at Stanford and co-founded Inception, a company developing language models based on diffusion.[1][3]
Start with Generative Modeling by Estimating Gradients of the Data Distribution →
Education
Links
X Stanford profile Co-founder · Inception
Associate Professor, Computer Science · Stanford University
Background
Ermon’s historical Cornell page records electrical-engineering degrees from Padova in 2006 and 2008 and doctoral work with Carla Gomes and Bart Selman. Stanford records the subsequent computer-science Ph.D. and his faculty appointment. His research spans generative modeling and computational sustainability, where inference methods are applied to problems involving uncertain data and decisions.[1][6][9]
Career
Contributions
From noise to samples
With Yang Song, Ermon developed score-based generative modeling: a model estimates which direction increases the probability of data, then uses those estimates to turn noise into samples. This offers a way to generate images without directly normalizing a probability distribution.[2]
Learning from preferred answers
Ermon also co-authored Direct Preference Optimization, or DPO. A common way to improve an assistant is to collect pairs of answers and ask which one a person prefers. Earlier training pipelines first learned a separate reward model, then used reinforcement learning to optimize the assistant against it. DPO derives a training objective that learns directly from the preference pairs. The paper’s contribution is a simpler route to the same constrained preference-learning problem, with experiments on sentiment control, summarization and dialogue.[7]
Ideas
Inception takes the diffusion approach into text generation. Conventional autoregressive models extend an answer one piece at a time; Inception describes Mercury as generating many pieces in parallel and refining the result. The intended payoff is lower response latency for interactive applications such as coding and voice. Mercury is the company’s team product. Ermon’s verified connection is as a co-founder, alongside his separately documented research authorship.[3][8]
Tools, writing & talks
Writing, research & talks
Start here
Learns gradients of the data distribution and uses Langevin dynamics to generate samples.
Start with the 2019 paper for the mathematical basis of score-based generation. It assumes probability, gradients and neural-network training.
Co-author[2]Stefano Ermon explains why Inception bets on diffusion language models over autoregressive ones, how Mercury reaches coding competitiveness at much higher tokens per second, and what that speed changes for latency-sensitive agent workloads.
Catalog workOriginal sourceStudies changes that make score-based model training and sampling work better.
Co-author[4]- Direct Preference Optimization: Your Language Model is Secretly a Reward Modelproceedings.neurips.cc
Trains language models from preferred and rejected answers with a direct objective instead of a separate reward model and reinforcement-learning loop.
Co-author[7]