TMallGS: Scaling Unified Feature and Sequence Modeling for Generative E-commerce Search

As industrial search and ranking systems transition from traditional Deep Learning Recommendation Models (DLRM) to compute-intensive Transformer architectures, they face a 'compute wall' and performance saturation. The authors investigate how to adapt Large Language Model (LLM) scaling strategies to the unique, heterogeneous requirements of e-commerce search without losing the high-precision signals critical for ranking.

The authors propose TmallGS, a ranking architecture that moves away from the 'all-in-one' tokenization common in LLMs. Instead, it employs a 'semantic divide-and-conquer' strategy consisting of five key innovations:

• Hierarchical Tokenization: A two-stage pipeline (Field-wise Saliency Reweighting and Distribution-Calibrated Projection) that maps diverse search features into optimized subspaces.
• Field-Adaptive Transformer: A backbone using per-field QKV projections and noise-adaptive gating to refine semantic interactions and suppress irrelevant historical noise.
• Decoupled FiLM Late Fusion: A mechanism that uses Feature-wise Linear Modulation (FiLM) to inject explicit high-frequency cross-features into the final representation, preventing the 'low-pass filter' effect of deep attention.
• Context-Aware Bias Net: A module that uses a global context anchor to orthogonally decouple systemic biases (like pagination) from genuine user intent.
• Error-Aware Progressive Training: A curriculum learning strategy that uses dynamically weighted losses to stabilize training in deep networks.