The LinkedIn algorithm papers everyone quotes (one was withdrawn)

More than most people realise. LinkedIn's engineering teams publish papers on arXiv and write up ranking changes on their own engineering blog. Three sources do most of the work in any honest explanation of the 2026 feed: the 360Brew paper, the LiRank paper, and LinkedIn's posts on dwell time.

This matters because almost every article about the LinkedIn algorithm is a summary of another article about the LinkedIn algorithm. Numbers get copied, rounded, and re-attributed until nobody can point at where they came from. The primary sources are public and free, so there is no good reason to work from the echo.

• 360Brew (arXiv 2501.16450), submitted January 2025: a single large model that handles ranking and recommendation across the platform.
• LiRank (arXiv 2402.06859), submitted February 2024: the production ranking framework behind feed, jobs and ads, and the place dwell time appears as an explicit objective.
• LinkedIn's Feed engineering blog, which has covered dwell time twice, in May 2020 and again in October 2024.
• From Features to Transformers (arXiv 2502.03417), submitted February 2025 and withdrawn by its authors in February 2026. This one is still being quoted.

From Features to Transformers: Redefining Ranking for Scalable Impact, arXiv 2502.03417, authored by a LinkedIn team led by Fedor Borisyuk. It was submitted on 5 February 2025 and reached a third version marked withdrawn on 8 February 2026.

The withdrawal notice on the arXiv listing is one sentence, and it is unusually direct.

We found discrepancies in the claims of the paper upon further investigation and therefore request to withdraw this submission from arXiv.

arXiv 2502.03417, withdrawal comment, version 3

The paper described a transformer-based ranking system, sometimes referred to as LiGR, and made three claims that were widely repeated: that it deprecated most hand-designed feature engineering and beat the previous system using only a handful of features rather than hundreds, that it validated scaling laws for ranking systems, and that scoring items together as a set improved diversity automatically.

The authors do not say which claims were wrong. Withdrawal covers the whole paper, so the honest position is that none of it should be cited as established, not that some specific number was off.

This is a normal and healthy thing for researchers to do. The problem is downstream: the paper sat up for a year, was summarised into dozens of marketing blog posts, and those summaries did not get withdrawn. If you have read that LinkedIn's ranking now uses almost no engineered features, or that set-wise scoring is why your feed feels more varied, that claim traces back here.

That LinkedIn built a 150-billion-parameter decoder-only model, on top of a Mixtral 8x22 mixture-of-experts base, capable of handling over 30 predictive tasks that each previously had their own dedicated model maintained over several years.

The architectural point worth understanding is that it reads. Rather than consuming hundreds of thousands of handcrafted numeric features, the paper describes processing all input as text: member behaviour, post content and profiles. It states that this eliminates the need for feature engineering and for maintaining complex graphs of model dependencies.

It was trained on LinkedIn's first-party data from non-EU members over about nine months, and tested at context lengths up to 88,000 tokens, with performance improving as more of a member's history was included. The paper also reports a larger advantage over the older system for members with fewer interactions, which is the cold-start case.

If you want that in practical terms rather than architectural terms, we wrote it up separately in how 360Brew ranks your feed.

LiRank describes the production ranking framework behind feed, job recommendations and ad click-through prediction. Its reported A/B results are modest and specific: a 0.5% lift in feed member sessions, 1.76% more qualified job applications, and a 4.3% improvement in ads click-through rate.

Those numbers are worth holding onto, because they are the scale at which real ranking improvements land. Any blog post promising that one formatting trick will multiply your reach is describing something larger than LinkedIn's own engineers report from rebuilding the ranking stack.

The part that matters for writers is how dwell is modelled. LiRank operationalises it as a binary long-dwell classifier: rather than predicting the exact seconds, it predicts whether a member's dwell will pass a context-dependent threshold, and uses that prediction directly as a feed ranking objective. LinkedIn's October 2024 engineering post covers why, and reports that dwell-based features increased the model's area under the ROC curve by as much as 10% offline, with a large decrease in skipped updates in live experiments.

We pulled the practical version of this into what dwell time means for your posts.

Anything about your specific post. None of these papers publish a ranking formula, a weight for saves against comments, an optimal post length, or a best time to publish. They describe architecture and report aggregate experiment results.

So when you read that saves carry five times the weight of a like, or that posts holding attention past 61 seconds see engagement rates above 15%, understand where those come from. They are not LinkedIn's published figures. They are inferences from third-party samples, mostly Richard van der Blom's Algorithm Insights work, which analysed roughly 1.8 million posts across about 400,000 profiles, and from tool vendors analysing their own users' data.

Third-party samples are useful. They are just not the same kind of evidence as a paper the platform published about its own system, and they should never be quoted as though they were.

That distinction is not pedantry. It is the difference between a claim you can check and a number that has been through six blogs. We went looking at how well the most-repeated engagement benchmarks hold up, and several of them contradict themselves inside a single article.

Three things follow from the primary sources without any inference. Write on a consistent topic, write things worth staying on, and stop optimising for the reaction rather than the read.

1. 1.Topic consistency compounds. A model that reads your posts as language and infers what you are expert in can only build that picture if your posts are about something. Scattered subject matter gives it nothing to learn.
2. 2.Depth beats hooks. If the ranking objective includes whether someone stays past a threshold, an opening line that wins the click but loses the reader is working against you.
3. 3.Write for the cold reader. 360Brew reports its largest gains over the older system for members with sparse interaction history, which means your post is increasingly being judged on whether it makes sense to someone who does not know you.
4. 4.Ignore the tactic of the week. LinkedIn's own engineers report single-digit-percent lifts from serious infrastructure work. Nobody's formatting trick beats that.

None of that is a growth hack. The system reads what you wrote and decides whether it was worth someone's time, so the work is the writing. If you want help getting the thing in your head onto the page in your own voice, rather than in the flattened register everyone's drafts default to, that is what innernote does. It interviews you, learns how you actually talk, and drafts from your material instead of from the internet's average opinion.