G-CAST: Governed Continual Self-Training for E-Commerce Agents under Evolving Demands
G-CAST is deployed in Alibaba’s 1688 procurement agent, with experiments reporting better adaptation and retention of existing behavior than an autonomous post-training baseline.
The system turns daily production evidence into training data as shopping requests and the agent’s tools change. It routes data using rewards and prioritizes it by entropy, while also governing changes to the rewards themselves. Each model update must pass cumulative evaluation to control the loss of previously accepted capabilities. Ablation experiments attribute the gains to both the data orchestration and governed reward evolution.
HF Daily Papers' note
The system turns daily production evidence into training data as shopping requests and the agent’s tools change. It routes data using rewards and prioritizes it by entropy, while also governing changes to the rewards themselves. Each model update must pass cumulative evaluation to control the loss of previously accepted capabilities. Ablation experiments attribute the gains to both the data orchestration and governed reward evolution.
HF Daily Papers' note
score 4