Tmall's TMallGS model scales Transformer search with 5 key innovations

In a paper submitted to arXiv, researchers from Tmall (Alibaba) propose TMallGS, a unified architecture for generative e-commerce search that scales Transformer-based ranking to handle heterogeneous features. Traditional deep learning recommendation models are giving way to compute-intensive Transformers. However, current approaches like OneTrans and Climber tokenize all features uniformly, ignoring their diverse nature. TMallGS addresses this with five components: (1) Hierarchical Distribution-Calibrated Tokenization (with Field-wise Saliency Reweighting and Distribution-Calibrated Projection) maps features into optimized subspaces; (2) a Field-Adaptive Gated Transformer backbone applies per-field QKV projections and noise-adaptive gating for better semantic interaction; (3) Decoupled FiLM Late Fusion preserves explicit high-frequency signals; (4) a Context-Aware Bias Net decouples systemic bias from user intent; and (5) Error-Aware Progressive Training uses dynamically weighted losses for robust learning.

Extensive offline experiments and online A/B tests on Tmall Search demonstrate that TMallGS significantly improves training throughput and achieves substantial gains in key metrics like UCTCVR (user click-through conversion rate) and GMV (gross merchandise value). The work aligns with the industry trend of applying scaling laws to ranking models, showing that careful architectural design can unlock predictable performance improvements. For e-commerce platforms, this means more relevant product searches, higher conversion rates, and better user experiences — all while keeping training costs manageable.