Meta's ConnectionMind fuses social graphs with LLMs to boost watch time by 0.43%

Recommendation systems on platforms like Meta have long struggled to model the complex web of social relationships – friendships, group memberships, and creator interactions – alongside massive heterogeneous content like text and video. Traditional models often ignore these signals or treat them independently, lacking the reasoning capability to integrate multi-relational context for fine-grained personalization. To address this, Meta researchers introduce ConnectionMind, a framework that tightly couples social network structure with large language models (LLMs) to achieve scalable, interpretable, and reasoning-aware recommendations.

ConnectionMind constructs a heterogeneous graph linking users, items, friends, groups, and creator pages. It then frames recommendation as a graph reasoning problem: discovering personalized paths from users to candidate items. An LLM-based policy reasons over these graph structures to guide decisions. To train at scale, the team first uses supervised fine-tuning (SFT) on large-scale user-item interaction trajectories, then applies end-to-end reinforcement learning (RL) to refine the model's social graph reasoning. Extensive experiments on real-world datasets show ConnectionMind outperforms representative baselines, and it has been deployed in Meta's large-scale recommendation pipeline. Online A/B tests show a 0.43% increase in video watch time, demonstrating measurable real-world impact from fusing social graphs with LLMs.