G-CAST
A governed continual self-training framework for e-commerce agents. It performs data orchestration through reward routing and entropy-prioritized sorting, and combines governed reward evolution with cumulative-evaluation-based checkpoint switching to achieve automated, safe continual model updates amid requirement drift and toolchain evolution.
Timeline (last 90 days)
The Alibaba 1688 team proposed the G-CAST system to address the continual learning problem of e-commerce conversational shopping agents under demand drift and evolving execution frameworks.
G-CAST converts everyday production data into training data routed by reward and prioritized by entropy, paired with a governed reward evolution mechanism.
G-CAST adopts governed checkpoint switching based on cumulative evaluation, where each model update must pass cumulative evaluation before going live.
The paper conducted experiments on a real production-grade e-commerce agent, and the results show that compared with a strong autonomous post-training baseline, G-CAST adapts better to newly introduced requirements while also better preserving previously accepted behaviors.
Ablation experiments show that both components—reward-routing plus entropy-prioritized data orchestration, and governed reward evolution—are important to the final gains, and removing either one reduces the effect.
G-CAST has been deployed on YuanCai (cai.1688.com), a B2B procurement agent on the Alibaba 1688 platform.
G-CAST is not an offline solution in a lab, but a continual training system running on real B2B procurement traffic.
All Facts (7)
The paper conducted experiments on a real production-grade e-commerce agent, and the results show that compared with a strong autonomous post-training baseline, G-CAST adapts better to newly introduced requirements while also better preserving previously accepted behaviors.
50%UnverifiedG-CAST converts everyday production data into training data routed by reward and prioritized by entropy, paired with a governed reward evolution mechanism.
50%UnverifiedG-CAST adopts governed checkpoint switching based on cumulative evaluation, where each model update must pass cumulative evaluation before going live.
50%UnverifiedThe Alibaba 1688 team proposed the G-CAST system to address the continual learning problem of e-commerce conversational shopping agents under demand drift and evolving execution frameworks.
50%UnverifiedAblation experiments show that both components—reward-routing plus entropy-prioritized data orchestration, and governed reward evolution—are important to the final gains, and removing either one reduces the effect.
50%UnverifiedG-CAST is not an offline solution in a lab, but a continual training system running on real B2B procurement traffic.
50%UnverifiedG-CAST has been deployed on YuanCai (cai.1688.com), a B2B procurement agent on the Alibaba 1688 platform.
50%