Yu et al. introduce G-CAST, a governed continual self-training system for LLM-powered e-commerce agents that addresses demand drift and harness evolution without periodic manual retraining. The system converts daily production evidence into training data through reward-routed, entropy-prioritized orchestration, selecting high-value samples under evolving demand and capability contracts. Governed reward evolution and checkpoint transition under cumulative evaluation control each model update, ensuring new behaviors are learned without sacrificing previously accepted capabilities. Deployed on Yuancai, a B2B procurement agent on Alibaba's 1688 platform, G-CAST demonstrates stronger adaptation to newly introduced demand while better preserving existing behavior compared to an autonomous post-training baseline. Ablation studies confirm that both the reward-routed, entropy-prioritized data orchestration and governed reward evolution components are necessary for these gains, validating the system's design for continuous adaptation in production e-commerce environments where user needs and tool capabilities evolve rapidly.

Read original