LinkedIn Feed ranking has undergone a significant overhaul, with the professional network deploying a new system powered by large language models and GPUs to serve more than 1.3 billion members. The company detailed the engineering work behind the update in a post authored by Hristo Danchev on March 12, 2026.
Why LinkedIn Rebuilt Its Feed Architecture
The previous system relied on multiple separate retrieval sources, including trending content indexes, collaborative filtering, and embedding-based pipelines. Each source maintained its own infrastructure, creating engineering complexity and making unified optimization difficult. Moreover, the old ranking model evaluated each post impression independently, missing the sequential patterns in how professionals consume content over time.
Unified Retrieval Through Fine-Tuned LLMs
LinkedIn replaced the multi-source architecture with a single artificial intelligence-driven retrieval system using LLM-generated embeddings. The system understands semantic relationships between topics rather than relying on keyword matching. For example, a member interested in electrical engineering who engages with posts about small modular reactors will receive relevant content, because the model connects those topics through world knowledge from its pretraining corpus.
The team also addressed a specific challenge with numerical features. Raw engagement counts passed directly to the model produced near-zero correlation (-0.004) between item popularity and embedding similarity scores. Converting those counts into percentile buckets wrapped in special tokens increased that correlation by 30 times. Furthermore, retrieval recall at the top 10 results improved by 15 percent as a direct result of the encoding change.
LinkedIn Feed Ranking With a Sequential Model
On the ranking side, LinkedIn built a Generative Recommender model that treats a member’s interaction history as an ordered sequence rather than a set of independent events. The model processes more than 1,000 historical interactions through a transformer architecture with causal attention. Consequently, it captures trajectory and long-term professional interests rather than scoring each post in isolation.
The model uses a Multi-gate Mixture-of-Experts prediction head with shared DCNv2 experts gated per task. Passive tasks such as clicks and long dwells receive specialized gating, while active tasks such as likes, comments, and shares share the same sequential representations. The company stated that the social platform audits the model regularly to ensure equitable treatment across its member base, relying on professional signals and engagement patterns rather than demographic attributes.
Training Efficiency and GPU Infrastructure
Training the sequential model on millions of member histories required substantial investment in computing infrastructure. A key finding was that filtering training data to include only posts that received positive engagement, rather than all impressed posts, reduced per-sequence memory footprint by 37 percent. In addition, the approach enabled processing 40 percent more training sequences per batch and delivered a 2.6 times faster training iteration on a production configuration using 8 H100 GPUs.
Adding just two hard negatives per member during training improved retrieval recall by 3.6 percent compared to using easy negatives only. The serving infrastructure uses a disaggregated architecture separating CPU-bound feature processing from GPU-heavy model inference. A custom Flash Attention variant called GRMIS delivered an additional 2 times speedup over PyTorch’s standard scaled dot-product attention implementation.
Impact on Member Experience
The combined system delivers content updates within minutes of new posts gaining traction, rather than hours. For newer members, the sequential model begins shaping the LinkedIn Feed from the first few interactions, reducing the cold-start disadvantage that affected earlier systems. The company stated that the changes represent a significant evolution in how it builds recommendation systems at scale, combining the semantic understanding of large language models with production-grade engineering.
“The Feed you see today is smarter, more personal, and more relevant than ever. These systems represent more than incremental improvements — they demonstrate significant evolution in how we approach building recommendation systems at scale.”
Hristo Danchev, LinkedIn Engineering