Feed SR is LinkedIn's Feed ranking model, which uses a member's history of posts and actions to predict how they will engage with new posts1. LinkedIn says its newer ranking considers profile information and engagement over time, and its engineering team says it never uses demographic attributes234. LinkedIn Engineering describes its Generative Recommender as processing more than a thousand past interactions and reports a 2.10% increase in time spent in online Feed tests45. Members' profile choices and engagement help shape recommendations, and LinkedIn says it is testing an interest picker during sign-up234. In March 2026, Tim Jurka described the Feed's aim as relevant, authentic conversations, and LinkedIn said it was reducing recycled posts and engagement bait; the sources do not establish the resulting effect on any individual creator's reach23.
Inputs, model size and test results
LinkedIn's Pressroom and Tim Jurka say the new ranking looks at what a member chose to share on their profile, such as industry, experience, skills and geography23. They also say the Generative Recommenders are augmented with large language models, which help the system understand what a post is actually about and how a member's interests change over time23. LinkedIn Engineering also says the ranking relies on professional information and patterns of engagement, never demographic attributes4. In size, LinkedIn Engineering's table lists the Generative Recommender stack at about 1.1 billion parameters, against about 500 million for the original DLRM stack5. LinkedIn Engineering reports that the Generative Recommender increased time spent by 2.10% in online Feed tests5.
What Feed SR is and how it ranks posts
From LinkedIn's research and patents, which may not describe how LinkedIn works today
LinkedIn's research paper from February 2026 describes the model's input as a sequence in which posts a member was shown alternate with the actions they took on them1. The model uses earlier events in that sequence when interpreting a later event1. The model is trained on member interaction histories to predict actions at every point in the sequence1. LinkedIn Engineering says earlier models assessed each post independently and produced a single predicted likelihood of an action, while its Generative Recommender (GR) processes more than a thousand past interactions and produces a separate prediction for each post in the sequence45. When ranking, the candidate posts are appended to the end of the sequence and scored all at once1.