Rec-Distill: An Industrial Distillation Pipeline for Large-Scale Recommendation Models
Published in arXiv preprint, 2026
Rec-Distill bridges the gap between large recommendation models and latency-constrained online serving. It combines large-teacher scaling with decoupled training, black-box distillation, debiasing, and a hybrid batch-streaming pipeline for dynamic recommendation environments. The framework scales teachers to 24B dense parameters and 20K behavior sequences while allowing lightweight students to recover more than 60% of teacher gains in the best setting, with improvements also validated in online recommendation and advertising scenarios.
Recommended citation: Haoran Ding, Wenlin Zhao, Yuchen Jiang, Juren Li, et al. (2026). Rec-Distill: An Industrial Distillation Pipeline for Large-Scale Recommendation Models. arXiv:2605.29755.
Download Paper