Contents
LLM-based ranking is LinkedIn's use of large language models (LLMs), AI systems that read text, to score (give a relevance rating to) and rank (put in order of relevance) feed posts, search results and creator recommendations1234. The models help understand what posts, queries, profiles and members' interests are about1234. The Feed uses information members choose to share on their profiles and their interactions over time to personalize posts from their network and people they are not connected to12. In October 2026, Tim Jurka explained that LinkedIn uses AI systems to recognize and reduce generic, repetitive content while also showing authentic conversations5.
Language models in Feed ranking and content quality
LinkedIn uses large language models (LLMs) in Feed ranking by adding them to larger models that read a member's activity over time to order posts, which it calls Generative Recommenders12. Tim Jurka wrote in March 2026 that the LLMs help the system better understand what a post is about and how a member's interests change over time2. LinkedIn's Pressroom says these systems look at details such as a member's industry, skills, experience and geography, alongside engagement over time1. Jurka wrote that the engagement details include what a member has interacted with and content the member has or has not reviewed2. In October 2026, Jurka wrote that the same AI systems that show authentic content also help identify and reduce generic, repetitive content, often called "AI slop"5. Larger "teacher" LLMs identify new slop patterns and produce training data, and that data trains smaller, efficient models that run at scale5. LinkedIn adapts these models with Low-Rank Adaption (LoRA) adapters, which adjust an existing model for a task, instead of building and serving a new model each time5.
Ranking creators to follow
From LinkedIn's research and patents, which may not describe how LinkedIn works today
For creator recommendations, LLMs help match member and creator profiles by turning them into numeric "meaning" fingerprints, called embeddings, to rank creators to follow, according to LinkedIn Engineering3. These fingerprints also help the later ranking model score possible creators by matching professional specialties, which interaction history and patterns among similar members struggle to capture, especially for members with little interaction history3. Training the ranker on the same embeddings keeps it from mis-scoring creators found this way relative to those found by methods using relationships between members or popularity3.
Prefill-only ranking in search
From LinkedIn's research and patents, which may not describe how LinkedIn works today
For search, LLMs directly score hundreds of results against each query, the member's profile and the job or profile content in AI Job Search and AI-powered People Search, according to LinkedIn's engineering team4. The team calls this prefill-only ranking because it uses only the model's initial understanding step, called the prefill, and skips text generation4.