Stefano Soatto

Career

Soatto received a PhD in control and dynamical systems from Caltech in 1996. Before joining UCLA in 2000, Soatto was a researcher at the University of Udine from 1995 to 1998, an assistant and then associate professor of electrical and biomedical engineering at Washington University, and a research associate in applied sciences at Harvard.²⁶

At AWS, a UCLA Samueli advisory-board biography describes Soatto as vice president of applied science for AWS AI, heading the AI Labs that supported services such as Amazon Comprehend, Kendra, Textract, Lex and Personalize. Amazon Science now lists Soatto as a vice president and distinguished scientist in the AWS Agentic AI organization, and UCLA lists Soatto as a computer science professor currently on leave.²¹³

Research

Soatto's UCLA homepage lists research interests in computer vision and nonlinear estimation and control, including shape and motion estimation, visual textures, shape from defocus, image analysis and statistical inference on manifolds, with an interest in how computers can use vision, sound and touch to interact with people and their surroundings. Work with Yi Ma, Jana Kosecka and Shankar Sastry on Euclidean reconstruction received the David Marr Prize, and work with Roger Brockett on optimal structure from motion received the Siemens Prize with an IEEE Computer Society Outstanding Paper Award.⁶⁷

Soatto's recent papers with AWS colleagues focus on language models and agents. Gated KalmaNet (CVPR 2026) introduces a memory layer based on test-time ridge regression, and EvoMAS (ICML 2026) evolves configurations of LLM-based multi-agent systems through feedback-driven mutation and crossover.⁴⁵¹

Work and contributions

• Gated KalmaNet: A Fading Memory Layer Through Test-Time Ridge Regression Published research

  Co-author (last author) of the paper, listed on Soatto's Amazon Science page as a CVPR 2026 publication. The abstract describes a Kalman-filter-based layer that keeps more of the past than fading-memory models and says the code and 8B and 32B hybrid models are released under Apache 2.0.⁴¹

• EvoMAS: Evolutionary Generation of Multi-Agent Systems Published research

  Co-author (last author) of the ICML 2026 paper, which evolves the configurations of LLM-based multi-agent systems through feedback-driven mutation and crossover.⁵¹

Timeline

1. 1996Received a PhD in control and dynamical systems from the California Institute of Technology.²⁶
2. 1999Received the David Marr Prize with Yi Ma, Jana Kosecka and Shankar Sastry.⁷⁶
3. 2000Joined the UCLA computer science faculty.²⁶
4. Nov 2025Co-authored "Gated KalmaNet", a fading memory layer for language models, as last author.⁴
5. Feb 2026Co-authored EvoMAS, on evolutionary generation of multi-agent systems, as last author.⁵

Other documented roles

Roles the sources document that are not confirmed as current.

• Assistant then Associate Professor of Electrical and Biomedical Engineering (before 2000), Washington University²⁶
• Researcher, Department of Mathematics and Computer Science (1995–1998), University of Udine²⁶

Honors and recognition

• ACM Fellow · Association for Computing Machinery, 2023⁷
• IEEE Fellow · IEEE, 2013⁷
• David Marr Prize · International Conference on Computer Vision, 1999⁷⁶
• Siemens Prize with the Outstanding Paper Award · IEEE Computer Society⁶
• NSF CAREER Award · National Science Foundation⁶

Where they work

• Amazon 39 people in the wiki
• UCLA 5 people in the wiki

Related people

• Vasi PhilominAlso at Amazon
• Sarthak JainAlso at Amazon
• Jitendra MalikAlso at Amazon
• Eric HartfordAlso at Amazon
• Clare LiguoriAlso at Amazon
• Winston WeinbergAgents · Language models
• Nick FrosstLanguage models · Computer vision
• Yuxin WuComputer vision · Language models
• Niki ParmarLanguage models · Computer vision
• Barret ZophLanguage models · Computer vision