Hamed Firooz is a principal staff AI scientist and director at Meta AI who builds foundation models for personalization, recommendation, and online safety. He led development of LinkedIn’s 360Brew, a 150-billion-parameter personalization model, and co-created Meta’s Hateful Memes Challenge, a benchmark for recognizing harmful meaning that emerges from combinations of images and text.
Firooz earned a doctorate in electrical engineering from the University of Washington in 2012. His early work included industrial monitoring at Falkonry, sales prediction and customer retention at Base CRM, and advertising relevance at LinkedIn. At Meta from 2018 to 2023, he worked on multimodal understanding and platform integrity, contributing to multimodal bitransformers for image-and-text classification and the Hateful Memes dataset and challenge, which supplied more than 10,000 examples designed to defeat shortcuts based on either modality alone. His work also included Facebook AI Multimodal, a framework for developing and deploying integrity models.
Returning to LinkedIn in 2023, Firooz led a team developing a foundation model for personalization. The resulting 360Brew research, co-authored with Maziar Sanjabi and colleagues, represented member histories, candidate content, and prediction tasks as text, allowing one 150-billion-parameter model to address more than 30 tasks without task-specific fine-tuning. In offline evaluations, it matched or exceeded specialized production models.
- Unified personalization models: Firooz designed 360Brew to transfer across recommendation surfaces, accommodate users with limited interaction histories, and support instruction-based personalization without rebuilding a separate model for every task.
- Progressive model distillation: His production approach transfers capabilities from large models through successively smaller ones, combining gradual pruning, sparse attention, and mixed-precision quantization to preserve ranking quality while improving serving efficiency.
- Precision-aware recommendation: Firooz keeps the final language-model output layer at higher numerical precision so closely ranked candidates remain distinguishable, and uses attention masks to score multiple items without letting candidates influence one another.
- Multimodal content understanding: The Hateful Memes work framed online abuse as a contextual problem: images and captions that appear innocuous separately can become harmful when interpreted together.
Firooz returned to Meta in August 2025. His subsequent research on reinforcement learning for content moderation investigates reward design, reasoning-oriented prompting, and label-efficient training for policy-grounded moderation systems.
