G-CAST: Governed Continual Self-Training for E-Commerce Agents under Evolving Demands
Summary
- Researchers introduce G-CAST, a governed continual self-training system designed for e-commerce agents dealing with evolving demands and changing system harnesses.
- G-CAST addresses demand drift and harness evolution by utilizing reward-routed, entropy-prioritized training data and governed reward evolution.
- Controlled checkpoint transitions via cumulative evaluation enable the agent to learn new behaviors without catastrophic forgetting of established capabilities.
- Real-world experiments demonstrate that G-CAST improves adaptation to new demands while preserving legacy behavior compared to standard autonomous post-training baselines.
- The system is deployed in Yuancai, a B2B procurement agent operating on Alibaba's 1688 platform.
Why it matters
G-CAST offers a production-tested approach to continually updating LLM agents under shifting user requirements and toolsets, preventing catastrophic forgetting while automating post-training.