LinkedIn 解读基于大语言模型的新内容推荐机制

An article from

The company’s Profile Relevance model uses LLM-generated embeddings to show users more relevant posts from people they don’t follow.

Published Nov. 9, 2026

By

This audio is auto-generated. Please let us know if you have feedback.

LinkedIn shared insights into its evolving content recommendation system, with an overview of its improved, artificial intelligence-powered process. The platform’s system determines what content from non-connections gets displayed to users in their feed.

In a post on the LinkedIn Engineering Blog, LinkedIn engineer Shiyong Lin explained the platform’s new Profile Relevance model, which uses LLM-generated embeddings to match members and posts, based on professional context.

The process essentially summarizes each user’s profile into a prompt, taking into account each member’s profile information and posting history. It then builds a personalized guide of what they’re likely to be interested in.

On a basic level, this means that the system creates a chatbot-style summary of each user’s information, which then dictates what they get shown in the app.  

Using these generated summaries, LinkedIn built a new recommendation system that measures content relevance versus each member’s summary.   

As per Lin: “The system evaluates how well each post matches a member’s professional context across specific factors, including topic, location, job title, company, industry, seniority, and education.”

LinkedIn assessed these results manually before building a system that showed users more relevant posts from members they don’t follow in the app.

This has led to notable improvements in feed relevance.

As per Lin: “Online experiments recorded a +0.54% lift in Feed Updates Viewed, a +0.94% increase in viral rate, and a -18.8% reduction in hide rate, indicating a more relevant and less noisy member experience. For members, this translates into a feed that surfaces high-quality content beyond their immediate network while remaining aligned with their professional context, even when they have limited historical activity.”

Lin said that this system will also benefit creators by ensuring that their posts will reach relevant audiences beyond their immediate network.

“Notably, for cold-start members with limited activity history, we observed a +17.7% increase in viral engagement and a +1.8% lift in DAU, underscoring the model’s ability to generalize effectively even with sparse interaction data.”

In summary, the post recommendations from non-connections in the LinkedIn feed are likely now more relevant and more valuable, driving a better LinkedIn experience.