WHALE unifies two complementary recommendation backbones: Wukong for high-order interactions among non-sequence features and HSTU for long user-behavior sequences. Each layer keeps both modules active and uses attention-based fusion, allowing Wukong-derived interaction representations to query HSTU-derived history representations repeatedly across the network. The authors also describe customized Triton kernels and model-systems co-design for industrial efficiency. The abstract reports consistent offline gains, positive online gains with a modest serving-throughput trade-off, and production deployment, but provides no numerical results.
No heat snapshots are available in the last 24 hours.