LinkedIn updates feed algorithm with LLM-powered ranking and retrieval

LinkedIn is launching a brand new AI-powered feed rating system that makes use of massive language fashions and GPUs to research publish content material and floor extra related updates to its 1.3 billion members.

Why we care. Understanding how LinkedIn surfaces content material is essential if you need your posts — or your model’s — to be found. The brand new system prioritizes topical relevance and engagement patterns, LinkedIn stated. Posts that reveal experience and align with rising skilled conversations could journey farther throughout the community — even with out current connections.

The main points. LinkedIn rebuilt a lot of its feed advice system utilizing massive language fashions, transformer fashions, and GPU infrastructure. The overhaul facilities on two methods: retrieving related posts and rating them within the feed.

Unified retrieval system. LinkedIn changed a number of separate discovery methods with a single LLM-powered retrieval mannequin.

• Beforehand, feed candidates got here from a number of sources, together with community exercise, trending posts, collaborative filtering, and topic-based methods.
• The brand new strategy makes use of LLM-generated embeddings to know what posts are about and the way they hook up with your skilled pursuits.
• Now, LinkedIn can hyperlink associated matters even after they use totally different terminology. For instance, engagement with posts about small modular reactors may floor content material about electrical grid infrastructure or renewable power.

Rating that follows your pursuits. After retrieval, LinkedIn ranks posts utilizing a transformer-based sequential mannequin. As a substitute of evaluating posts independently, the mannequin analyzes patterns throughout your previous interactions — together with likes, feedback, dwell time, and different indicators.

• This helps LinkedIn detect how your skilled pursuits evolve and suggest content material that displays these shifts.

System efficiency and infrastructure. The system runs on GPU infrastructure designed to course of thousands and thousands of posts whereas maintaining feeds recent.

• The structure can replace content material embeddings inside minutes and retrieve candidates in beneath 50 milliseconds, LinkedIn stated.

Enhancing feed high quality and authenticity. LinkedIn additionally announced updates to enhance content material high quality:

• Cracking down on automated engagement. LinkedIn is taking motion in opposition to remark automation instruments, browser extensions, and engagement pods that create inauthentic conversations. These instruments violate platform guidelines and undermine actual skilled discussions, LinkedIn stated.
• Lowering engagement bait and generic posts. LinkedIn plans to indicate much less content material designed purely to drive feedback or clicks — together with posts asking folks to remark “Sure” to spice up attain, posts pairing unrelated movies with textual content to sport distribution, and recycled thought-leadership with little substance.
• Serving to new members personalize their feeds sooner. LinkedIn is testing an “Curiosity Picker” throughout signup that lets new customers select matters equivalent to management, job search expertise, or profession progress, serving to ship related content material from day one.