Alibaba Deploys Self-Training Agent on 1688 Procurement Platform
TL;DR
- Alibaba has put G-CAST, a governed continual self-training system, into production inside Yuancai, a B2B procurement agent on the 1688 platform.
- The system routes daily production evidence into reward-routed, entropy-prioritized training data and gates each update through a cumulative-evaluation checkpoint rule.
- The abstract claims gains over a strong autonomous post-training baseline on both new-demand adaptation and prior-behavior preservation, but publishes no specific numbers.
Alibaba has deployed a self-training conversational agent on its 1688 B2B procurement platform, according to a paper posted to arXiv on November 13 by Yuechun Yu, Wenzheng Kong, Xiaofeng Mao, and Jin Zhang. The system is called G-CAST; the agent is called Yuancai.
The authors frame production as two kinds of drift. User demand shifts as new product categories and procurement scenarios appear, and the "product's execution harness-including its tools, skills, and operating procedures-continues to evolve." Periodic manual retraining, they write, is "too delayed and labor-intensive for this setting."
G-CAST's answer is to take daily production evidence from the live platform and funnel it into "reward-routed, entropy-prioritized training data and governed reward evolution." Each model update has to pass what the paper calls "Governed checkpoint transition under cumulative evaluation" before it is allowed to replace the running checkpoint, which the authors say lets the agent "learn newly introduced behaviors without sacrificing previously accepted capabilities."
The headline result in the abstract: G-CAST "achieves stronger adaptation to newly introduced demand while better preserving previously accepted behavior compared with a strong autonomous post-training baseline." The abstract names no specific accuracy numbers, no traffic volume on 1688, no deployment date, and no backbone model. Ablations, the authors add, show "both reward-routed, entropy-prioritized data orchestration and governed reward evolution are important to these gains."
G-CAST, the paper concludes, "is now deployed in Yuancai, a B2B procurement agent on Alibaba's 1688 platform."
Originally reported by paper
Read the original article →Original headline: G-CAST: Alibaba Ships First Governed Continual Self-Training Agent in Live B2B Procurement