Taobao's GALA recommender boosts orders 0.55% for 200M daily users

Alibaba researchers introduced GALA (Generative Aligned Learning for Adaptive Multimodal Representation), a production-ready recommender system for Taobao Shangou, the company's food delivery platform. Published on arXiv and accepted at ICDE 2026, GALA tackles a classic problem: how to effectively combine images, text, and user behavior signals without losing semantic alignment. Traditional two-stage systems pretrain image-text encoders separately from behavior-driven ranking models, causing a gap between what's semantically similar and what users actually click or order.

GALA's key innovation is an intermediate "generative RL alignment" stage. After behavior-aware triplet pretraining on query-image-text pairs, it applies GRPO (a reward-driven optimization technique) to refine multimodal embeddings using conversion-based rewards. This aligns representations with downstream objectives before a final adaptive gating stage merges multimodal and ID embeddings with a hybrid loss. The system is live in production, serving over 200 million daily active users. Offline benchmarks show consistent AUC gains of +0.12 and +0.20 over state-of-the-art baselines, and large-scale online A/B tests report a 0.55% increase in order volume—a significant lift for a platform of this scale. GALA demonstrates that bridging pretraining and ranking with RL-based alignment can deliver measurable business impact while remaining robust across diverse demand patterns.