Beyond Item IDs: Scaling Short-Form-Video Recommendation via Semantic-Native Long Sequence Modeling

Researchers present a production-deployed recommendation system that scales short-form video suggestions to billion-user scale by replacing traditional Video IDs with semantic-native representations and introducing a compression transformer to reduce computational complexity. The framework achieves order-of-magnitude improvements in memory efficiency and enables longer user behavior sequences, delivering measurable gains in user engagement and content consumption metrics.

This work addresses fundamental scalability challenges in recommendation systems that power platforms serving billions of users. Traditional approaches using orthogonal Video IDs create sparse semantic representations requiring massive embedding tables, while transformer-based sequence modeling faces quadratic computational costs that limit sequence length under production constraints. The solution combines two complementary innovations: semantic IDs leverage content structure through depth-truncated representations that generalize to new videos via shared prefixes, and a Global-Aware Compression Transformer uses temporal folding and global query integration to condense sequences without standard self-attention's quadratic overhead.

The significance lies in bridging the gap between academic recommendation research and industrial deployment realities. Most published work operates under unrealistic assumptions about computational budgets and latency constraints. This framework demonstrates that thoughtful representation design—moving beyond one-hot encodings to semantic hierarchies—can simultaneously improve model capacity and reduce infrastructure costs. The approach naturally handles cold-start problems through semantic prefix sharing, a persistent challenge in recommendation systems.

For the recommendation systems industry, this work validates that semantic representations outperform orthogonal embeddings at scale. Platform operators managing billions of user-video interactions face acute pressure to balance model depth with inference latency and hardware costs. The reported online A/B test improvements in user engagement metrics suggest meaningful business impact beyond engineering efficiency gains. The production deployment status indicates real-world validation, not theoretical promise, making this particularly valuable for engineers building large-scale systems.

• →Semantic IDs reduce embedding table size from corpus cardinality while improving cold-start generalization through shared semantic prefixes.
• →Global-Aware Compression Transformer achieves order-of-magnitude memory reduction and computational efficiency gains compared to standard transformers.
• →System successfully deployed at billion-user scale with measurable online improvements in user engagement and content consumption metrics.
• →Temporal folding and unified global query integration enable longer effective sequence lengths within strict production latency and resource constraints.
• →Approach bridges academic recommendation research and industrial deployment by solving real infrastructure cost and latency challenges.