How the LinkedIn feed algorithm works in 2026

LinkedIn publishes more about its feed than most platforms, and the newest papers are from 2026

LinkedIn’s engineers post papers about the feed on arXiv, the open research archive. The company also makes official statements in its pressroom and on its blog. Here’s what I read, newest first, and what each one says about whether it describes a live system.

• Feed SR (Hertel et al., LinkedIn, first posted February 2026, revised September 2026). The main feed ranker. The paper says it replaced the previous ranker and “has been serving the majority of LinkedIn’s Feed traffic.” confirmed · 1
• Connected Content Retriever (Gupta et al., LinkedIn, September 2026). How posts from your own network get narrowed down before ranking. The abstract reports online experiments on the live feed, but doesn’t say whether it’s fully launched. confirmed · 3
• Large Scale Retrieval for the LinkedIn Feed (Ramanujam et al., LinkedIn, October 2025). How posts from outside your network are found, with a fine-tuned version of Meta’s Llama 3 language model. It reports an online A/B test on members, not a full rollout; the 2026 ranking paper lists it among the feed’s retrieval sources. confirmed · 2, 1
• LiRank (Borisyuk et al., LinkedIn, 2024) and LiGR (Borisyuk et al., LinkedIn, 2025). Earlier generations of the feed ranker. Both papers say their methods went into production. confirmed · 4, 5
• LinkedIn Post Embeddings (Ramanujam et al., LinkedIn, 2024, updated 2025). The model that turns a post’s text into a representation of what it’s about. It’s used in feed retrieval and ranking, and the paper says it has been in production for over two years. confirmed · 6
• 360Brew (Firooz et al., LinkedIn, 2025). A 150-billion-parameter research model. Its own abstract calls it pre-production, and arXiv’s administrators have since withdrawn every version over a licence issue. confirmed · 7
• Official statements: LinkedIn’s March 2026 pressroom post on feed changes, and a May 2022 post on its official blog about feed quality. confirmed · 8, 9 Three older engineering blog posts, from 2017, 2018 and 2020, fill in details. They’re old, and systems change. confirmed · 10, 11, 12

What I couldn’t read: LinkedIn’s Help Center, its Professional Community Policies and its newest engineering blog post weren’t accessible when I checked. Nothing on this page relies on them.

A post reaches a feed in five steps, and LinkedIn has described each one

1Understanding the post. As a post goes live, LinkedIn’s classifiers label it “spam”, “low-quality” or “clear” in near real time. That’s how LinkedIn described it in 2017. confirmed · 12 A text model turns your words into a representation of the topic, so posts about similar things end up close together. confirmed · 6 To find readers outside your network, the post is also described to a language model as plain text: the type of post, what’s shared (text, image, video, job change), the author’s name, headline, company, industry and title, how popular the post is, the title and source of any linked article, and the text itself. confirmed · 2 A new post enters that search index within a minute. Its representation is refreshed within 30 minutes as people interact with it. confirmed · 2

2Finding candidates. There are two main pools. The first is your network: posts from people you’re connected to or follow. They make up over 70% of feed impressions and engagement. confirmed · 3 This pool also holds posts one of your connections reacted to, commented on or reshared without writing them. LinkedIn’s engineers call these “stranger viral”. confirmed · 3 That index holds over a billion items. Picking activity from your network narrows it to tens of thousands, and a pre-ranking model passes a few hundred to the ranker. confirmed · 3 The second pool is suggested content from outside your network. Here, a fine-tuned Llama 3 model reads a text description of you: headline, summary, industry, skills, location, work and education history, certifications, languages, and the posts you engaged with. It finds the closest posts among hundreds of millions and keeps the top one to two thousand. confirmed · 2 Every candidate has to pass LinkedIn’s trust classifiers, be in a language you understand, not come from someone you blocked, and not be something you’ve already seen. confirmed · 2

3Ranking by predicted actions. The ranker typically scores about 512 candidates per request. confirmed · 1 For each one, it predicts how likely you are to stay on the post past a set threshold (“long dwell”), like, comment, share or click. The 2024 paper also lists “vote”. confirmed · 1, 4 Those predictions are combined with weights into one score. confirmed · 1, 4 The 2026 model bases its predictions on your last 1,000 feed impressions and what you did on each, plus how close you are to the author (“viewer-to-author affinity”), how popular the post already is, and a language-model summary of your profile. confirmed · 1 Its engineers write that popularity signals “remain crucial”. confirmed · 1 In 2020, LinkedIn also described adding two things to a post’s score: the activity it would create downstream if you engaged, and the value to the author of getting feedback. confirmed · 10 In 2018, it described a change that moved about 8% of all feedback from the top 0.1% of creators to the bottom 98%. confirmed · 11 Both are years old, and today’s weights aren’t published.

4Final rules. After ranking, “additional business logic is applied on the ranked post list.” confirmed · 1 The 2025 LiGR paper says what that has historically included: at least two items between posts from outside your network, and at least two items between posts by the same author. confirmed · 5 The 2026 paper doesn’t list the rules in use now.

5Learning from what happens next. The ranking models are updated every day with new interaction data. confirmed · 1 Posts you were shown and didn’t engage with become “hard negatives” when the retriever is trained: examples of what not to fetch for you. confirmed · 2 In 2020, LinkedIn described a model that predicts whether you’ll skip a post and lowers the post’s score in proportion to that risk. confirmed · 10 LinkedIn also reports running “explore/exploit” methods in the feed. confirmed · 4 In general, that means sometimes showing items the model is less sure about, to learn from the response. inferred

What it means for what you post: hold attention, earn real replies, say what it’s about

These are my deductions from the confirmed pieces above. LinkedIn hasn’t published them as rules.

• Write for the read, not the click. The ranker predicts long dwell and has penalised likely skips, which suggests a post that holds the people who stop on it has more going for it than one that only earns a quick tap. inferred
• Your network is still the main road. Most impressions come from connections and follows, and a reaction, comment or reshare from a connection can drop your post into their network’s pool. So reach grows through people who actually engage. inferred And since the ranker weighs how close each viewer is to the author, the people who interact with you often are likely the ones who’ll see your next post. inferred
• Say what the post is about, in your audience’s words. The out-of-network retriever reads your text, headline, company, industry and title, and matches them against people who engaged with similar posts. Clear topic words and an accurate headline give it more to match on. inferred
• Early response is a signal, not a stopwatch. How popular a post already is feeds both retrieval and ranking, so real engagement in the first hours carries forward. No source describes a fixed window after which a post is done. inferred
• Don’t bait. LinkedIn says it’s reducing “comment to agree” prompts, recycled and click-driven posts, and videos that don’t match their text, and that it’s working to make engagement pods ineffective. confirmed · 8 Posts people scroll past also become training examples of what not to show them. confirmed · 2
• Space your posts out. Given the same-author gap rule, posting several times in a row probably won’t fill back-to-back slots in anyone’s feed. No source gives an ideal posting frequency. inferred

What nobody outside LinkedIn knows, me included, and six beliefs checked

The biggest unknown is the weights. No published paper or official page says how much a comment counts against a like, a share or a long read. I look at posts all day, and I still can’t see those weights. Nobody outside LinkedIn can. The long-dwell threshold isn’t published either: the 2024 paper gives the 90th percentile of time spent as an example, and the 2026 paper says it depends on the post type. confirmed · 4, 1 The 2026 paper doesn’t list every post feature its ranker uses. Here are six common beliefs, checked against the sources (see the myths page).

• “External links are punished.” No evidence either way. No paper or official page I checked mentions a link penalty. The retrieval model reads a linked article’s title and source as part of the post. confirmed · 2 LinkedIn says it’s reducing “click-driven posts”, but it doesn’t define the term or mention links. confirmed · 8
• “Hashtags boost reach.” No evidence. The only mention of hashtags in the papers I read is as a training label that helps LinkedIn’s text model learn what posts are about. confirmed · 6 That suggests they’re read like any other topic words. Nothing shows they add reach. inferred
• “The golden hour decides everything.” No evidence for a fixed window. Popularity counts are inputs, and the retriever refreshes them within 30 minutes. confirmed · 1, 2 The 2017 description of watching how fast likes, shares and comments arrive was about spotting posts that might go viral, so low-quality ones could be reviewed. confirmed · 12
• “Comments beat likes.” Not confirmed. Each action is predicted separately and weighted, but the weights are unpublished. confirmed · 1, 4 A reaction, a comment or a reshare by a connection can each bring a post into their network’s pool. confirmed · 3
• “Dwell time counts.” Confirmed. Long dwell is a predicted action in the 2024 and 2026 rankers. confirmed · 4, 1
• “LinkedIn’s feed now runs on 360Brew.” Not supported. 360Brew was described as pre-production, and its paper was withdrawn. confirmed · 7 The 2026 production paper describes a different model, and says an LLM-based ranker it tested “never achieved superior online performance” over production. confirmed · 1 Language models are used elsewhere: to retrieve suggested posts and to summarise member profiles. confirmed · 2, 1

Company Pages: none of the sources I checked describes separate treatment for Page posts. The FAQ below says what can be inferred.