LinkedIn's AI Powers Smarter Follows

LinkedIn is overhauling how members discover creators, moving beyond popularity metrics to understand genuine interests. The professional network’s new recommendation engine, detailed by LinkedIn Engineering, uses Large Language Models (LLMs) to create a shared understanding of both users and creators.

This shift aims to surface relevant content and experts more effectively, particularly for new users who previously received sparse, popularity-driven suggestions. The system now places members and creators into a unified embedding space, allowing matches based on topical alignment rather than just follower counts.

From Profiles to Embeddings

The core of the new system involves transforming semi-structured profile data, like bios, headlines, and skills, into natural language prompts. These prompts are then fed into a fine-tuned LLM, which generates vector embeddings. These embeddings represent the latent interests and expertise of individuals.

Narrative prompting was chosen over templated approaches for its ability to handle missing fields and capture nuanced semantics. This approach aligns well with LLMs' training on vast natural language corpora, enabling deeper inference of user interests.

Fine-Tuning for Connection

To optimize these embeddings for follow prediction, LinkedIn employed supervised contrastive learning. A bi-encoder architecture, using shared LLM parameters, processes both member and creator prompts separately. The model is trained to pull embeddings of users and creators who are likely to connect closer together in the vector space, while pushing dissimilar pairs apart.

This fine-tuning process utilizes techniques like Parameter-Efficient Fine-Tuning (PEFT) and mixed-precision training to manage computational costs at scale. Gradient checkpointing further aids memory efficiency for these large models.

Retrieval and Ranking, Reimagined

Two embedding-based retrieval (EBR) pipelines, offline and online, were developed. The offline pipeline handles large-scale, batch candidate generation using Ray for distributed GPU inference and FAISS for efficient vector search. This is suitable for regular updates where latency is less critical.

For real-time personalization, especially during member onboarding, an online EBR pipeline provides low-latency retrieval. The fine-tuned LLM encoder runs on LinkedIn's internal Proxima infrastructure, generating embeddings on-the-fly for immediate use by a hosted vector search service.

Embeddings as Features

Beyond candidate generation, these rich embeddings are integrated as features into downstream ranking models. They help bridge the gap left by traditional behavioral and collaborative filtering methods, especially for cold-start users. To manage the high dimensionality of LLM embeddings, a task-aware supervised projection reduces their size before they are concatenated with other features.

This approach ensures that candidates from all sources, legacy generators and the new LLM-based EBR, are scored consistently within the ranking model. The projection layers are then extracted for efficient online inference.

Why This Matters for Professional Networks

This move by LinkedIn highlights a broader industry trend: the increasing reliance on LLM-based semantic understanding to move beyond shallow metrics. Competitors in the professional networking and talent acquisition space, such as The Bot Company (StartupHub score 64/100), Apptronik (score 59/100), and Agility Robotics (score 56/100), are also exploring sophisticated AI to better match needs and opportunities. While LinkedIn itself has a StartupHub score of 17/100, its strategic implementation of LLMs for core product features demonstrates its ongoing commitment to AI innovation. The company has also reportedly raised $1B (Series B, 2025), indicating significant investment in its technological future.

The implications are far-reaching. For end-users, it means a more personalized and valuable experience, uncovering niche experts and relevant content they might otherwise miss. For creators, it promises better audience reach and engagement. For the platform itself, deeper user understanding translates to increased retention and activity.

The Road Ahead

LinkedIn identifies two key areas for future enhancement: richer prompts incorporating real-time engagement signals beyond static profile data, and more expressive matching techniques, potentially moving beyond simple cosine similarity to capture finer-grained relationships. These advancements build upon the foundational LLM-based semantic retrieval system now in place.