Meta turns to generative AI to revamp its recommendation feeds

Meta is rethinking how recommendation engines work from the ground up.

Running the feed for billions of users across Facebook, Instagram, and Threads usually requires massive dense retrieval clusters. Every time you open an app, traditional systems calculate your embedding and compare it against billions of items in a giant vector store. As catalogs grow, storage and compute costs skyrocket.

Now, two new research papers from Meta show a completely different approach: treating recommendations as a next-token prediction problem.

Predicting items instead of searching databases

Instead of matching vectors across a massive database, generative retrieval systems treat your interaction history as a prompt.

The system first assigns each item a Semantic ID (SID) that captures its context and properties, using architectures similar to TIGER. An encoder creates the unique SID based on the item description.

Then, a transformer model looks at your past sequence of SIDs and predicts the next SID in the chain. Because the model predicts the ID directly rather than searching the whole catalog, inference and storage costs stay constant even as the catalog expands. Engineers can even tweak generation temperature to adjust recommendation diversity.

Fixing cold starts with LIGER and Mender

Pure generative retrieval has one major flaw: cold starts. It overfits to items seen during training, making it terrible at handling newly added items or fresh users with no history.

To bridge this gap, Meta developed LIGER, a hybrid architecture. LIGER trains on both similarity scores and next-token objectives. During inference, it uses generative retrieval to pick candidates, injects cold-start items, and ranks the final set using candidate embeddings.

In a second paper, Meta introduced Mender, a multimodal preference discerner. Mender uses an LLM to parse user actions, like reading praises or complaints in reviews, into explicit category preferences. The core recommendation model then conditions on both your history and these extracted preferences to generate the next item.

Why it matters

For consumers, recommendation feeds should feel significantly more intuitive. Instead of trapping you in repetitive clickbait loops, models using techniques like Mender can actually pick up on nuanced feedback and adjust recommendations accordingly.

For the tech industry, this signals a massive shift in recommendation infrastructure. Vector databases and nearest-neighbor search have dominated the last decade of ad and feed architecture. If generative retrieval proves it can keep compute costs flat as catalogs scale, it will force e-commerce giants and streaming services to rewrite their retrieval stacks.

We think Meta's hybrid approach with LIGER is the smartest play here. Pure end-to-end generative retrieval sounds great in theory, but ignoring the cold-start problem is commercial suicide for ad and marketplace platforms. By keeping dense retrieval as a fallback, Meta gets the compute savings of generative models without breaking its core monetization loops.