Maziar Sanjabi is a machine-learning researcher at Amazon AGI who helped develop 360Brew, LinkedIn’s foundation model for personalized recommendations, and FedProx, a framework for training machine-learning models across heterogeneous devices. His research spans optimization, federated learning, recommendation systems, and the conditions under which large language models generalize.
Sanjabi earned a doctorate and master’s degree in electrical engineering at the University of Minnesota and completed postdoctoral research at UCLA and the University of Southern California. Earlier roles included research at Starkey Hearing Technologies and work on deep generative models for computer graphics and game design at Electronic Arts. At Meta, he researched scalable and responsible AI systems, including multimodal learning, explainability, privacy, and harmful-content detection.
He coauthored FedProx, which adapts federated optimization to participants with different datasets and computational resources. His related work on federated multi-task learning and fair resource allocation examines how distributed systems can accommodate distinct local objectives without concentrating benefits among better-resourced participants.
As a principal scientist at LinkedIn AI in 2025, Sanjabi helped build 360Brew, a unified model that interprets member profiles, interaction histories, candidate items, and task instructions together. He described the system as supporting more than 30 ranking and prediction tasks while reducing specialized feature engineering.
Distinctive research contributions
- Personalization through in-context learning: Member histories and instructions become model inputs, enabling recommendations for unfamiliar tasks and people with limited prior activity.
- Train large, then distill for production: Sanjabi described training Brew-XL, a roughly 150-billion-parameter model, before transferring its capabilities into smaller, production-ready systems; starting with a small model produced weaker results. His conference presentation also identified training data, model size, and usable history as separate levers for recommendation quality.
- Longer context requires genuine generalization: Additional user history improved recommendations until the model encountered context lengths it could not reliably handle, separating the value of information from the model’s ability to exploit it.
- Excessive chain-of-thought can weaken transfer: In research on reasoning demonstrations, Sanjabi and collaborators found that overexposure to explicit reasoning traces can impair generalization when those traces disappear; mixing examples with and without reasoning improved transfer.
