LinkedIn's engineering team detailed a ground-up rebuild of the feed — a single LLM-based pipeline that matches posts to members by meaning and can surface relevant content from people you don't follow.
2026-11-10 · by Moe Ameen
On March 12, 2026, LinkedIn's engineering team published "Engineering the next generation of LinkedIn's Feed," a detailed post by Hristo Danchev describing a ground-up rebuild of the system that decides which posts show up in your feed. The old setup stitched together several specialized retrieval sources — trending content, collaborative filtering, and older embedding systems — each tuned separately. The new one replaces that fragmented stack with a single pipeline built on large language models.
The system has two stages. First, retrieval: a dual encoder built on a fine-tuned, distilled LLM turns both "member prompts" and "post prompts" into embeddings in a shared space and matches them by cosine similarity, so a post is retrieved because it means something close to what a member is interested in — not because of keyword overlap or who they follow. A GPU-based nearest-neighbor search returns candidates in under 50 milliseconds from an index of millions of posts. LinkedIn also converts raw engagement counts into percentile buckets wrapped in special tokens, which it says sharply improved how well a post's popularity lines up with its embedding.
Second, ranking: the shortlist is ordered by a transformer LinkedIn calls the Generative Recommender. It treats a member's feed interaction history as a sequence of more than 1,000 past interactions, adds count and affinity features after the transformer (late fusion), and uses a mixture-of-experts head to score passive and active engagement separately. (This Generative Recommender is distinct from 360Brew, an earlier LinkedIn research model that some coverage has conflated with the live feed ranker.)
LinkedIn lists several things it says the rebuild changes for distribution: posts from people you don't follow can reach you when they match your interests semantically; new members can get relevant content from their profile before they've engaged much; fresh posts are embedded within minutes, so timely content can surface faster; and the feed adapts within a session as interests shift. LinkedIn also says it audits the system so posts from different creators "compete on equal footing" and that ranking uses professional signals rather than demographic attributes. These are LinkedIn's own claims — the post includes no independent evaluation — and the underlying work is described in two research papers on causal-LM retrieval and sequential feed ranking.
The deep shift here is from who-you-know to what-you're-about. Because the retriever matches posts to people by meaning, the creators who win are the ones who publish a steady stream of genuinely on-topic content around a clear set of subjects — and who show up everywhere that same idea can land, not just in one feed. Relevance-based reach means your best idea isn't capped by follower count, which is exactly the kind of leverage a generation engine is built to press.
That is where [Kompozy](/) fits. Point it at a source — a talk, a client call, a hot take — and it generates a week of LinkedIn-ready material governed by your [Persona Brief](/glossary/persona-brief): text posts kept tightly on-subject so the dual encoder has something clean to embed, brand-exact [Carousel Posts](/glossary/output-buckets) and quote graphics via [HyperFrames](/glossary/hyperframes), a long-form article for the days depth travels, and an avatar [Persona Short](/glossary/persona-shorts) for video. You review each one behind the per-post gate, schedule into your best windows, and let [Autopilot](/glossary/autopilot) publish — not only to LinkedIn but across the eight social platforms plus blog and email from one queue. A recommendation system that rewards clear, consistent subject matter rewards exactly what Kompozy is built to produce at volume, and spreading the same on-topic ideas across every destination means no single platform's model is the only thing deciding whether your work gets seen.
In a March 12, 2026 engineering post, LinkedIn described replacing a stack of separate retrieval models with a single LLM-based pipeline. A dual-encoder LLM matches posts to members by meaning (retrieval), and a transformer it calls the Generative Recommender orders the results by reading your interaction history as a sequence (ranking).
It is the ranking model in LinkedIn's rebuilt feed. The transformer treats a member's feed interaction history as a sequence of more than 1,000 past interactions and uses a mixture-of-experts head to score passive and active engagement separately. It is distinct from 360Brew, an earlier LinkedIn research model that some coverage has conflated with the live feed ranker.
According to LinkedIn, yes — more so than before. Because retrieval matches posts to members by semantic relevance rather than only network ties, a post can reach interested members who don't follow you when it matches their interests. LinkedIn frames this as content competing "on equal footing" regardless of the creator, though it offers no independent evaluation.
Less than it did, by LinkedIn's description. The new system leans on what a post is about and a member's inferred interests rather than purely on who follows whom, which loosens the follower-count ceiling for on-topic content. Followers still help, but relevance is doing more of the work in deciding reach.