Most people assume their LinkedIn feed is shaped by everything they do on the platform. Every like, every scroll, every “I don’t want to see this” click. That all of it feeds some giant behavioral model that decides what to show them next.
It doesn’t.
The architectural choice nobody talks about
LinkedIn’s retrieval system, a fine-tuned LLaMA-3 model with 3 billion parameters, builds a 3,072-dimensional embedding of every member on the platform. That embedding is what determines which content even enters the candidate pool for your feed. If a post doesn’t match your embedding closely enough, the ranking engine never sees it. It’s not deprioritized. It doesn’t exist.
Here’s the part that matters: that embedding is built from your profile information and your positive engagement history only. Likes, comments, shares. That’s it.
LinkedIn tested including negative signals. Dismissals. “Don’t show me this” clicks. Hide actions. The result? Retrieval quality got worse.
They shipped the positive-only version.
This is a detail I picked up reading Trust Insights’ Unofficial LinkedIn Algorithm Guide, Q1 2026, and it’s the kind of architectural decision that sounds minor but changes how you should think about the entire platform.
What the system actually sees
Your LinkedIn identity, at least to the retrieval engine, is a composite of two things: the words on your profile and the content you actively approved of. Not what you scrolled past. Not what you marked as irrelevant. Not the posts that annoyed you enough to hit “hide.”
None of that registers.
The system learns who you are exclusively from what you endorsed with an action. Your retrieval embedding is a portrait painted only in things you reached toward.
This means the common instinct to “train” your feed by dismissing irrelevant posts is, at the retrieval level, doing nothing. The dismiss button might affect other parts of the platform, but the engine that decides which posts are even eligible for your feed? It never sees the dismissal. It only knows what you liked.
The asymmetry that matters
There’s a deeper implication here for anyone creating content or trying to build an audience on LinkedIn.
Your readers’ embeddings are shaped only by their positive actions. When someone likes your post, that interaction enters the text prompt that generates their member embedding. Your content’s language, your concepts, your vocabulary become part of how LinkedIn understands that person. A like on a post about demand generation nudges their embedding toward the demand generation neighborhood.
But when someone scrolls past your post? Nothing happens. Their embedding stays exactly where it was.
This creates a compounding asymmetry. People who engage with your content become increasingly likely to see more of it, because every positive interaction pulls their embedding closer to your content’s semantic position. People who ignore it don’t drift away. They just stay still. The gravitational pull only works in one direction.
What this changes about feed curation
The practical upshot is counterintuitive. Most people think of feed curation as a two-lever system: amplify what you want, suppress what you don’t. On LinkedIn’s retrieval layer, only one lever is connected to anything.
That makes your positive engagement choices carry unusual weight. Every like, every comment, every share is a vote for where you want to exist in LinkedIn’s semantic space. Not in some vague algorithmic sense. Literally. Your engagement enters the text prompt that a 3-billion-parameter language model uses to generate your coordinates in a 3,072-dimensional vector space.
The corollary: passive scrolling is invisible. Extended periods of consuming content without engaging don’t degrade your embedding. They don’t improve it either. They’re a null input. The system simply doesn’t update its understanding of you until you do something affirmative.
For content creators, this reframes the engagement question entirely. The goal isn’t to avoid negative reactions. It’s to generate positive ones. A post that 90% of people scroll past and 10% of people actively engage with is, to the retrieval system, performing exactly as well as a post that 90% of people enjoyed but didn’t act on and 10% engaged with. The silent majority is literally invisible.
The design philosophy underneath
There’s something worth sitting with in this design choice. LinkedIn’s engineers tried the version that learns from rejection, and it made the system worse at understanding people.
Negative signals are noisy. A person hides a post for a dozen different reasons: wrong timing, too long, bad mood, already read something similar, don’t like the author’s tone. The signal is real but diffuse. It tells the system what someone didn’t want in one specific moment without clarifying what they do want.
Positive signals are clean. A like is a like. A comment is a deliberate investment of attention. The signal points in a clear direction.
In a system that needs to place you precisely in a space with a billion other members, clean signal wins. Even if it means throwing out half the available data.