ShopX: Alibaba researchers built one model to handle agentic shopping from intent to item

A team of researchers from Alibaba (affiliated with Taobao) has published ShopX, a foundation model designed to bridge the gap between natural language intent and item-space fulfillment in AI-powered shopping agents. Traditional approaches wrap an LLM around existing search and recommendation APIs, forcing complex intents through limited retrieval or ranking interfaces—creating lossy hand-offs. ShopX instead treats the entire process as a single model-native task: it takes a user's intents, plans execution steps, and directly operates on item-space using LLM-operable semantic IDs (SIDs). This allows ShopX to perform operations like SID beam-search retrieval, listwise ranking, and product bundling without needing external tools.

To build ShopX, the team designed semantically recoverable SIDs and a training recipe that equips a general LLM for multi-turn, flexible item-space fulfillment while retaining general knowledge and instruction-following capabilities. They evaluated ShopX against standard tool-mediated agentic systems on single- and multi-turn fulfillment tasks derived from anonymized Taobao production logs. Results show that model-native fulfillment improves overall framework behavior, particularly on complex or ambiguous requests. The paper (arXiv:2606.31693) positions ShopX as a step toward truly intent-driven agentic shopping experiences that move beyond page- and feed-based browsing.