What is it about?
Online shopping platforms must quickly select useful products from extremely large catalogues. However, conventional recommendation systems usually pass products through several separate stages, allowing errors to accumulate and often requiring a different model for each shopping setting. OxygenREC provides a unified alternative. Before a customer requests a recommendation, a large language model uses recent searches, browsing behaviour, shopping context, and general knowledge to produce a short description of the customer’s likely intent. When a recommendation is needed, a smaller and faster model combines this description with relevant past actions to generate suitable product candidates. The same system can support different stages of the shopping journey, from homepage discovery and product feeds to checkout. It also learns to balance the different goals of these settings and to search for better candidates. Tests on JD.com’s production traffic showed improvements in clicks, conversions, order volume, and the total value of goods ordered, while maintaining response times of 50–80 milliseconds.
Featured Image
Photo by Julio Lopez on Unsplash
Why is it important?
OxygenREC demonstrates how the reasoning ability of large language models can be used in a real-world recommendation system without placing a costly language model directly in the real-time serving path. By preparing compact intent instructions in advance and using an efficient model when a request arrives, the framework combines richer intent understanding with the speed required by large e-commerce platforms. Its unified design also reduces the need to build and maintain separate recommendation models for different shopping scenarios. In production A/B tests across six settings at JD.com, OxygenREC increased order volume by 1.26% to 9.14% and gross merchandise value by 1.64% to 10.99%, compared with existing production strategies. The deployed system supports tens of thousands of queries per second, indicating that instruction-guided generative recommendation can deliver measurable business value at industrial scale.
Perspectives
Our starting point was that effective recommendation requires more than predicting which product a customer may click. A system should also consider why an action occurred and what the customer may need in the current shopping context. This led us to separate the recommendation process into “slow thinking” and “fast action”: richer reasoning is performed before the request arrives, while a lightweight model responds immediately when the recommendation is needed. In this design, the large language model acts as an intent interpreter rather than the direct online recommender. Compact instructions connect its reasoning ability with efficient product generation, while multi-scenario optimization allows one model to serve different stages of the shopping journey. More broadly, our work suggests that practical AI systems do not always have to choose between sophisticated reasoning and real-time efficiency; they can distribute these tasks across different time scales.
Zhiwei Zhang
Read the Original
This page is a summary of: OxygenREC: An Instruction-Following Generative Framework for E-commerce Recommendation, July 2026, ACM (Association for Computing Machinery),
DOI: 10.1145/3805712.3808404.
You can read the full text:
Contributors
The following have contributed to this page
