LinkedIn is updating the feed algorithm with LLM-powered ranking and retrieval

LinkedIn is launching a new AI-powered feed system that uses large-scale language models and GPUs to analyze post content and surface updates that are most relevant to its 1.3 billion members.

Why do we care. Understanding how LinkedIn presents content is critical if you want your posts – or your product – to be found. The new system prioritizes topic relevance and engagement patterns, LinkedIn said. Posts that show expertise and are relevant to the discussions of emerging professionals can go far in the network – even without existing connections.

Details. LinkedIn rebuilt most of its feed recommendation system using large-scale language models, transformer models, and GPU infrastructure. The fix focuses on two programs: retrieving related posts and ranking them in the feed.

Integrated retrieval system. LinkedIn has replaced several separate recruiting systems with a single recruiting model powered by LLM.

• Previously, feed candidates came from many sources, including network activity, trending posts, collaborative filtering, and topic-based systems.
• The new approach uses the embeds generated by LLM to understand what the post is about and how it connects to your professional interests.
• Now, LinkedIn can link related topics even when they use different words. For example, interacting with posts about small reactors can generate content about electric grid infrastructure or renewable energy.

A degree that follows your interests. After retrieval, LinkedIn ranks posts using a transformer-based ranking model. Instead of examining posts independently, the model analyzes patterns in all of your previous interactions – including likes, comments, dwell time, and other signals.

• This helps LinkedIn see how your interests are evolving and recommend content that reflects those changes.

System performance and infrastructure. The system runs on a GPU infrastructure designed to process millions of posts while keeping the feed fresh.

• The architecture can review embedded content within minutes and retrieve candidates in less than 50 milliseconds, LinkedIn said.

Improving feed quality and authenticity. LinkedIn also announced updates to improve content quality:

• Reducing automatic engagement. LinkedIn is taking action against comment automation tools, browser extensions, and engagement pods that create fake conversations. These tools violate the platform’s rules and undermine genuine job interviews, LinkedIn said.
• Reducing engagement bait and general posts. LinkedIn plans to show minimal content designed solely to drive comments or clicks — including posts that ask people to comment “Yes” to increase reach, posts that pair unrelated videos with game distribution, and recycled thought leadership with minimal content.
• Helping new members personalize their feed quickly. LinkedIn is testing an “Interest Selector” during registration that allows new users to select topics such as leadership, job search skills, or career growth, helping to deliver relevant content from day one.