Generative AI Advances for Intelligent Commerce: Alibaba's Knowledge‑Driven, Reasoning, and Creative Technologies

2023 has become the year of generative AI large models. Since the launch of ChatGPT, a wave of AI technologies has rapidly spread worldwide, prompting industry leaders, startups, and research institutions to release dozens of general‑purpose and domain‑specific models.

At the 2023 ACM China Turing Conference SIGAI China Forum, Alibaba’s Alimama CTO Zheng Bo presented the vision of “Intelligent Commerce 2.0”, emphasizing three defining traits: knowledge‑driven operations, logical reasoning, and creative intelligence.

Knowledge‑driven innovations include a multimodal product detection system for the “Paizhi” feature, which aligns text prompts with images to improve detection mAP by 2.1 %. A two‑stage pre‑training pipeline (billions of e‑commerce image‑text pairs followed by supervised fine‑tuning on hundreds of millions of transaction pairs) yields a unified multimodal representation that achieves >98 % top‑100 recall for same‑item retrieval. The team also built a large‑scale multimodal training platform (MDL + AiLake) that can train 5 billion samples on 100 A100 GPUs within two days, and launched the AI Serving4LM engine with trillion‑parameter model serving capabilities.

Logical reasoning is embodied in the AIGA (AI Generated Action) decision‑intelligence framework. It integrates RL‑based bidding, learning‑based auction design, and novel algorithms such as Deep GSP, Neural Auction, and Two‑stage Auction. The AIGB (AI Generated Bidding) model reframes bidding as a conditional generation problem, outperforming traditional RL baselines on public datasets. Data‑driven decision pipelines combine LLM‑based intent understanding, OLAP‑driven multidimensional analysis, and AI analyst‑generated data stories to democratize knowledge for millions of small merchants.

Creative intelligence covers AI‑generated marketing creatives, virtual try‑on models, and AI‑styled fonts. The system uses diffusion models to synthesize product backgrounds, predicts optimal text placement, and generates context‑aware copy. Virtual models are created via a multi‑stage generation process with texture‑control networks, enabling fully customizable avatars for fashion merchants. Font style transfer, trained on historic stone‑inscriptions, yields five free commercial fonts and was presented at CVPR 2023.

Beyond product applications, Alimama’s “Wanxiang” platform leverages the above technologies to provide end‑to‑end AI‑powered marketing solutions, from user acquisition to large‑scale promotions.

Academically, Alimama collaborates with leading universities through the PAAI (Peking‑Alibaba AI Innovation) Joint Lab, producing over five papers accepted at top conferences (KDD, IJCAI, WWW) and advancing large‑scale graph models, decision intelligence, and AI‑generated music.

For more information, visit the lab website: http://paai.pku.edu.cn/