← back
360Brew: LLM-based Personalized Ranking and Recommendation - Hamed and Maziar, LinkedIn AI
Takeaway
One large LLM trained on promptified user behavior and distilled small can replace many specialized recommenders with strong cold-start gains.
Summary
- LinkedIn's 360Brew replaces many task-specific recommenders with one LLM foundation model that ranks across feed, jobs, search, etc. with zero-shot, in-context, and instruction-following capabilities.
- 'Promptification' converts user profile + interaction history + candidate item into a prompt; the LLM predicts relevance instead of a classical retrieval+rank pipeline.
- Train Brew-XL (150B params, Mixtral architecture upcycled) for max quality, then distill down to ~3B for production serving.
- Three scaling levers: more training data, larger model size, and longer context (more user history) all improve performance — but model generalization caps long-context gains.
- Largest wins are on cold-start users (<5 interactions) where world knowledge in the LLM beats data-hungry baselines.
recommendationllmlinkedin
Original description
We will give a talk about our journey of building a foundation model for solving ranking and recommendation tasks About Hamed Firooz Principal AI Scientist at LinkedIn Core AI. With 15 years in large-scale AI, Hamed leads the 50-person team behind LinkedIn’s 150-billion-parameter foundation model that personalizes the experience for hundreds of millions of members. Before LinkedIn, he led multimodal Content Understanding model at Meta AI that handle tens of billions of daily requests. His work spans open-source projects like Hateful Memes benchmark dataset and papers at venues such as NeurIPS and ICML. About Maziar Sanjabi Maziar is a Principal Scientist at LinkedIn AI, where he leads efforts in training large language models (LLMs) for personalization tasks. Prior to joining LinkedIn AI, he worked at Meta AI, applying AI research to the development of multimodal systems for real-world applications. With over a decade of experience in AI research across both industry and academia, Maziar has a proven track record of building and scaling cutting-edge AI technologies, including LLMs, multimodal systems, and privacy-aware AI. He has published over 60 papers, many of which have been featured in top-tier AI conferences such as NeurIPS, ICML, ICLR, ACL, EMNLP, and CVPR. Recorded at the AI Engineer World's Fair in San Francisco. Stay up to date on our upcoming events and content by joining our newsletter here: https://www.ai.engineer/newsletter