Meta News

For years, recommendation systems have largely been built around the same fundamental idea: collect signals about a user, generate a set of candidate items, score those candidates individually and rank them according to the probability that the user will click, watch, buy or otherwise engage with them.

Large language models are beginning to challenge that architecture.

Instead of treating every candidate as an isolated prediction problem, companies are experimenting with recommendation systems that understand a user’s activity as a sequence. The conceptual shift is similar to what happened in natural-language processing: rather than predicting whether an individual word is relevant, an LLM learns to predict what should come next given everything that came before.

⁠Software Engineering Daily's discussion with LinkedIn VP of Engineering Tim Jurka explores how this approach is already being applied at enormous scale. LinkedIn recently rebuilt significant parts of its feed around sequence modeling, treating recommendations increasingly like the next-token prediction problem underlying modern LLMs.

Traditional recommendation systems typically operate through several stages. First, candidate generation reduces millions of possible items to a manageable subset. Ranking models then calculate scores for those candidates using signals such as previous clicks, connections, interests, popularity and predicted engagement.

The models are extremely effective, but they often evaluate candidates independently.

Sequence models introduce a different question.

Instead of asking, “How likely is this user to interact with this post?” the system can ask something closer to:“Given everything this user has recently seen and done, what content should logically come next?”

That resembles the fundamental training objective behind an LLM. A language model receives a sequence of tokens and predicts the next token. A recommendation model can receive a sequence of interactions and predict the next interaction.

LinkedIn’s feed redesign applies this concept to content recommendation. The company combines sequence modeling with traditional recommendation signals rather than simply replacing its existing infrastructure with a chatbot-style LLM.

The distinction matters. The value of LLM-inspired architectures is not necessarily that recommendations become conversational. It is that techniques developed for understanding sequences of language can also be applied to sequences of human behavior.

Consider someone who reads several posts about Kubernetes, then watches a video about cloud infrastructure and subsequently follows an engineer discussing distributed systems.

A conventional recommendation model might independently determine that the user has an interest in Kubernetes, cloud computing and software engineering.

A sequence model can potentially capture something more subtle: the relationship between those actions and how the user’s interests are evolving.

Order matters.

Reading about databases after spending a week investigating AI infrastructure may mean something different from reading exactly the same database article after researching frontend development.

This is analogous to language. Words derive much of their meaning from context and position. The same principle can apply to behavior.

That makes recommendation increasingly resemble a problem of modeling a user’s trajectory, rather than maintaining a static profile describing what the user likes.

There is another major advantage.

Traditional recommendation systems depend heavily on structured signals: clicks, likes, shares, follows, categories and manually engineered features. LLMs can also reason about the semantic meaning of the content being recommended.

A model does not necessarily need thousands of previous interactions to understand that two articles discuss similar technologies. It can infer relationships from the content itself.

That capability can improve one of the persistent problems in recommendation systems: understanding new content.

A newly published article has little engagement history. Traditional collaborative filtering therefore has limited evidence for deciding who should see it. A language model can immediately analyze the article’s subject, technical depth, tone and relationship to other content.

This creates an opportunity to combine behavioral and semantic understanding.

The system knows both what the content means and how the user’s interests have evolved.

One of the more interesting consequences discussed in the Software Engineering Daily episode is the possibility of using natural language to influence recommendation quality.

Recommendation platforms constantly make editorial decisions that cannot easily be represented by a single engagement metric. A platform might want to discourage repetitive content, reduce low-quality posts, encourage professional discussions or avoid recommendations that technically generate clicks but create a poor user experience.

Historically, those objectives have required combinations of classifiers, manually engineered rules and ranking adjustments.

LLMs offer another mechanism: policies expressed in natural language.

Because the model understands language, engineers can potentially describe characteristics of desirable or undesirable content more directly and use those policies as another signal within the recommendation architecture. LinkedIn is exploring this combination of LLM capabilities and traditional signals as part of its feed engineering.

This could make recommendation systems more controllable, but it also introduces new evaluation challenges. Natural-language policies are inherently less deterministic than conventional rules.

The architecture becomes much harder when deployed to hundreds of millions of users.

LLMs are computationally expensive compared with traditional ranking models. A recommendation service cannot simply run a massive frontier model against every possible post whenever someone scrolls their feed.

Recommendation systems operate under strict latency requirements. Users expect the next items to appear almost instantly, while platforms may need to evaluate enormous numbers of potential recommendations.

Inference cost therefore becomes an architectural constraint.

LinkedIn’s work involves balancing these newer modeling techniques with existing recommendation signals and infrastructure while managing inference costs at massive scale.

This is likely to become the dominant engineering pattern across the industry. LLMs will not necessarily replace every ranking model. Instead, different models will operate at different stages depending on cost and complexity.

Smaller models can perform inexpensive filtering. Traditional systems can generate candidates. More sophisticated models can then reason over a dramatically reduced set of possibilities.

The challenge becomes deciding where expensive intelligence produces enough improvement to justify its cost.

Recommendation systems have traditionally relied heavily on measurable signals such as click-through rate, watch time, conversion and engagement.

LLMs introduce objectives that are harder to quantify.

A recommendation may generate fewer clicks while being substantially more useful. Another may produce immediate engagement but gradually make the feed repetitive. A third may improve diversity without producing an obvious short-term metric increase.

This creates a fundamental product problem.

Optimizing recommendation purely for engagement is relatively straightforward mathematically. Optimizing for something like “useful, professionally relevant and diverse content that doesn’t feel repetitive” is considerably more difficult.

LLMs may help systems reason about those concepts, but companies still need reliable ways of measuring whether the models are actually improving them.

That makes evaluation one of the central engineering challenges of LLM-based recommendation systems, alongside inference cost and scalability.

The longer-term implication is larger than simply producing better ranking algorithms.

Recommendation engines and generative AI are beginning to converge.

A traditional recommendation system selects something that already exists. A generative model can understand the user’s intent, retrieve relevant information, summarize multiple sources and potentially generate a response specifically for that individual.

The boundary between search, recommendation and generation therefore starts to blur.

A future feed might not simply decide which article should appear next. It could determine that several articles collectively answer something the user appears to be investigating and generate a personalized explanation connecting them.

At that point, recommendation becomes less about ranking objects and more about constructing an information experience.

LinkedIn’s experiment is important because it shows that the architecture behind LLMs has applications extending well beyond chatbots.

Human behavior itself can be represented as a sequence.

A click is an event. A viewed post is another. A follow, comment or skipped recommendation provides additional context. Taken together, those events form something resembling a behavioral language.

The engineering question then becomes remarkably similar to the one that transformed natural-language processing:

Given everything that came before, what should come next?

Recommendation systems have spent decades becoming better at predicting what people might like. LLM-inspired architectures suggest the next generation may instead become better at understanding where a user’s interests are going.

That is a considerably more ambitious problem—and potentially a much more powerful recommendation system.