Meituan's NONTP boosts recommendation hit rate by 34%

Next-Token Prediction (NTP) is widely used in generative recommendation but suffers from two structural limitations. First, it only optimizes single-step prediction, missing long-range behavioral structure (temporal locality). Second, in multi-domain sequences, each target embedding is updated only from the immediate previous hidden state, ignoring cross-domain context (spatial locality). To overcome these, the Meituan team introduces NONTP, which adds two lightweight auxiliary objectives during training. Temporal Contrastive Learning (TCL) uses a BYOL-style EMA teacher and InfoNCE loss to align hidden states with a K-step future trajectory. Trans-Domain Learning (TDL) mean-pools cross-domain hidden states and predicts through the shared head, creating a second gradient path without extra parameters. Both are discarded at inference, ensuring zero overhead.

Results are striking. On a four-domain Meituan industrial full-ranking dataset, NONTP achieves HR@10 +34.3% over NTP and +18.3% over the state-of-the-art MBGR. On the public Amazon Movie-Book-CDs benchmark, it delivers HR@10 +2.8% and NDCG@10 +3.7%. Real-world A/B tests confirm a 1.8% lift in click-through rate and a 2.1% lift in gross merchandise value (both p<0.01). Ablation studies verify each component contributes independently, and gradient conflict analysis points to future refinements. The work highlights how simple, targeted signal augmentation can significantly improve generative recommenders without adding inference complexity.