Meta's SlimPer slims down AI personalization models for faster, deeper recommendations

Meta's new SlimPer model tackles a fundamental mismatch in recommendation systems: Transformer-style architectures, borrowed from generative AI, treat each user interaction as a separate token and scale compute with sequence length. But recommendations only need a single relevance score per <user, item> pair. SlimPer instead reformulates ranking as iterative refinement of a compact, unified knowledge base. At each layer, it selectively queries raw multi-modal user tokens, computes explicit matching scores, and updates the knowledge base—all with a fixed-size intermediate representation and O(N) per-layer cost. This decouples model depth from user history length, enabling deeper relevance understanding without proportional growth in compute or memory. The model also shares a single copy of user-side tokens across all candidate items (request-only optimization), and unifies sparse, dense, and sequence features within one backbone while providing interpretable attention mechanisms.

SlimPer has been deployed on Instagram Reels and Feed, two of Meta's largest recommendation surfaces. The researchers report measurable improvements in user engagement alongside a streamlined system architecture. By handling over 10,000 fine-grained user history events effectively, SlimPer demonstrates that personalization models can be both slimmer and smarter—achieving better performance with lower computational overhead. This work represents a practical breakthrough for industrial-scale recommendation systems, where managing long user histories under strict latency budgets is a critical challenge.