Profile Relevance Boosting Feed Relevance With Llms At Linkedin

LinkedIn has described how it uses a language model to find posts for members from people they are not connected to. Feeds usually lean on who you follow and what you and similar people have clicked, which, as LinkedIn's engineers write, makes posts from outside your network hard to surface and hurts creators still building an audience. Their system, Profile Relevance, matches on what the profile and the post say instead. It builds a text summary of each member's profile from the headline, summary, skills, jobs and education, leaving out name and gender, and runs it through an LLM embedding model, which turns text into a list of numbers so that texts about similar things land close together. Posts get the same treatment from their text, titles, transcripts and the author's location, and the system pulls posts whose numbers sit near a member's. It is one of several sources feeding the redesigned feed described in an earlier LinkedIn post.

Because nothing here is trained on clicks, LinkedIn had no click data to grade it with. Engineers first read about 20 member and post pairs per model variant by hand, then built AutoEval, in which an LLM judges whether a post fits a member's topic, location, job title, company, industry, seniority and education. After tuning its instructions until it agreed with human reviewers, they could grade 1,000 pairs per variant, and LinkedIn says Profile Relevance scored a substantially higher on-topic rate than its previous retrieval source. The same tool caught a weakness: a version that first had a large LLM condense profiles and posts into topics lost location cues, so the team now appends member and author country names to the text.

Topical match alone was not enough to earn a place in the feed. LinkedIn found that well-matched posts struggled to compete with posts other members had already engaged with, so a post from this source now needs at least one like, comment or share from someone outside the reader's network before it is recommended. With both changes, LinkedIn reports small overall gains in its online experiments, 0.54% more feed updates viewed and an 18.8% drop in posts hidden, and larger ones for members with little activity history, including 17.7% more viral engagement (LinkedIn's term for likes, comments and shares) and 1.8% more daily active users. Those figures come from LinkedIn's own tests, without stated sample sizes or durations. They suggest the profile text gives LinkedIn something to recommend from for members whose click history is too thin to rely on, although the system still waits for an engagement signal before showing a post.