Semantic IDs and efficient transformers scale video recommendations to billion-user sequences

Capturing user interests across extensive watch histories is critical for short-form video recommendation, but scaling sequence length has been bottlenecked by the semantic sparsity of atomic Video IDs and the quadratic complexity of Transformers. A team of researchers (affiliated with a major video platform) presents a production-deployed framework that addresses both issues. They replace orthogonal Video IDs with content-native Semantic IDs, using depth-truncated coarse-grained IDs to drastically shrink embedding table size and naturally handle cold-start content via shared semantic prefixes.

To overcome the sequence scaling barrier, they introduce a Global-Aware Compression Transformer that leverages non-parametric temporal folding and unified global query integration. This approach condenses the sequence, reducing peak memory footprint by an order of magnitude and drastically cutting computational overhead. The efficiency gain enables supporting longer sequence lengths at affordable cost in production, yielding substantial online improvements in satisfied user engagement and content consumption in large-scale A/B tests. The work has been accepted at SIGIR 2026.