G-CAST: Enabling E-commerce Agents to Keep Evolving Amid Demand Drift

Alibaba 1688 proposes G-CAST, using governed continual self-training to let e-commerce Agents learn new demands daily without losing old behaviors.
The Alibaba 1688 team proposes G-CAST, a governed continual self-training system for e-commerce conversational Agents facing demand drift and execution framework evolution. It converts daily production data into reward-routed, entropy-prioritized training samples, uses governed reward evolution, and switches checkpoints based on cumulative evaluation to learn new behaviors without losing accepted capabilities. Experiments on a production-grade Agent show better adaptation and retention than a strong autonomous post-training baseline, with ablations confirming both components matter. It is deployed in 1688's YuanCai B2B procurement Agent.
Conversational shopping in e-commerce is becoming an important application scenario for large-model Agents, but a repeatedly validated challenge is this: user demands drift, and product toolchains also upgrade—a model trained once quickly becomes inadequate. The G-CAST system proposed by the Alibaba 1688 team attempts to resolve this contradiction with "governed continual self-training"—enabling the Agent to automatically learn new capabilities from daily production data while not losing the old behaviors users already rely on.

Two real-world problems that make models "expire"
The paper points out that e-commerce Agents face two types of continuous change. The first is demand drift: new product categories, procurement scenarios, and feature requests keep emerging, and the distribution, intent, and complexity of user requests are all changing; a one-off trained model will gradually fall behind real traffic. The second is execution framework evolution: the tools, skills, and operational workflows behind the Agent continue to increase, and the model must learn to correctly invoke newly introduced capabilities without breaking existing interaction patterns that users have already become accustomed to.
These two types of change compound, making "periodic manual retraining" both slow and expensive. Manual retraining latency is measured in weeks or months, while demand drift happens every day. G-CAST's starting point is to automate this loop and add governance mechanisms.
How G-CAST works: from production evidence to governed model updates
G-CAST's core idea is to turn daily production evidence into training signals. Specifically, it converts daily production data under evolving demand and capability contracts into training data that is routed by reward and prioritized by entropy, paired with a governed reward evolution mechanism.
Two designs are worth noting. The first is data orchestration: not all production logs are equally important. G-CAST routes samples using reward signals and then prioritizes them by entropy, concentrating training compute on samples with greater information content. The second is reward evolution: as new demands and new tools emerge, the reward function itself also needs adjustment, but the adjustment must be controlled; otherwise the model may sacrifice old capabilities in pursuit of new rewards.
In the model update stage, G-CAST adopts governed checkpoint switching based on cumulative evaluation. Every update must pass cumulative evaluation before going live, ensuring that the new model acquires new behaviors while not losing previously accepted capabilities.
Reward routing and entropy-based prioritization are two key techniques in data orchestration. Reward routing means assigning training data to different processing paths based on sample scores under the reward model (for example, high-reward samples are used for reinforcement, while low-reward samples are used for filtering or correction), thereby preventing low-quality samples from contaminating training. Entropy-based prioritization uses the model's uncertainty about sample predictions (entropy) to select samples with greater information content, prioritizing training on cases the model is currently "unsure" about, improving learning efficiency. Combined, the two can quickly locate the most valuable training signals in massive production logs and reduce wasted compute.
Cumulative evaluation is a safety verification mechanism before a model goes live. It does not only look at the immediate effect of a single update; instead, it comprehensively compares the new model's performance on a historically accumulated evaluation set (including both old and new tasks). Only when the new model is no worse than the old model and shows improvement on new demands is checkpoint switching allowed. This mechanism can effectively prevent "catastrophic forgetting"—that is, the model suddenly losing old capabilities while learning new ones. In low-fault-tolerance scenarios such as B2B procurement, cumulative evaluation acts as a quality gate for automatic updates.
Experimental conclusions: adapt to new demands while preserving old behaviors
The paper conducted experiments on a real production-grade e-commerce Agent. The results show that, compared with a strong autonomous post-training baseline, G-CAST adapts better to newly introduced demands while also better preserving previously accepted behaviors. Ablation experiments further show that both components—reward-routed plus entropy-prioritized data orchestration and governed reward evolution—are important to the final gains; removing either one reduces the effect.
This system is currently deployed in YuanCai (cai.1688.com), a B2B procurement Agent on the Alibaba 1688 platform. Judging from the paper's description, G-CAST is not an offline solution in a laboratory, but a continual training system running on real B2B procurement traffic.
Ablation experiments are a commonly used validation method in machine learning papers. By removing key components of the system one by one and observing performance changes, they determine each component's contribution. G-CAST's ablation experiments separately removed reward-routed plus entropy-prioritized data orchestration and governed reward evolution. The results show that neither can be omitted, indicating that data orchestration is responsible for "learning efficiently" and reward evolution for "learning safely"; together they support the closed loop of continual self-training.
Implications for e-commerce Agent engineering
The value of G-CAST does not lie in proposing some new model architecture, but in turning "continual learning" into a governable engineering closed loop. Competition in e-commerce Agents is to a large extent competition in data flywheels: whoever can faster convert production traffic into model capability while controlling regression risk can maintain responsiveness amid changing demands. Governed checkpoint switching and reward evolution are essentially safety valves added to automatic updates—especially critical in low-fault-tolerance scenarios such as B2B procurement.
Another signal worth noting is that the paper discusses "toolchain evolution" and "demand drift" as equally important. Most continual learning work focuses only on changes in data distribution, whereas in Agent scenarios the capability contracts themselves are also changing, and the model needs to adapt simultaneously to changes in the input distribution and the available action space. This perspective has direct reference value for teams building Agent products.
One-line Summary
Alibaba 1688 proposes G-CAST, using governed continual self-training to let e-commerce Agents automatically learn new demands every day without losing old behaviors.
Paragraph Summary
The Alibaba 1688 team proposes the G-CAST system, which implements governed continual self-training to address two major problems facing conversational e-commerce Agents: demand drift and execution framework evolution. The system converts daily production data into training samples routed by reward and prioritized by entropy, paired with a governed reward evolution mechanism, and uses cumulative-evaluation-based checkpoint switching to ensure the model learns new behaviors without losing accepted capabilities. Experiments on a real production-grade e-commerce Agent show that G-CAST outperforms a strong autonomous post-training baseline in both adapting to new demands and preserving old behaviors, and ablations confirm that both data orchestration and reward evolution components are important. The system has been deployed in 1688's B2B procurement Agent "YuanCai."
Related articles

RePlan-VLA: Progress-Aware Replanning Breakthrough for Long-Horizon Robotic Manipulation
The RePlan-VLA framework converts VLA models' short-horizon capabilities into reliable long-horizon execution through progress-aware adaptive replanning, improving success rates by 21.9% and 10.4% on RoboCasa365 and CALVIN, and validating real-world deployment on a Franka Research 3 robot.

Geometric Bridge Flow Matching: A New Breakthrough in Blind Rigid Motion Correction
Geometric bridge flow matching enables blind rigid motion correction in head CBCT, achieving 37.1 dB PSNR and 0.27 mm RPE on simulated data in ~9.1 minutes, outperforming existing methods.

Google Releases Gemini 4 Hydrogen: 2.5x Faster Without Sacrificing Intelligence
Google releases the lightweight Gemini 4 Hydrogen: 2.5x faster task completion, AA Intelligence Index of 54 beating its flagship sibling, 7.3s single-step on Agent Arena faster than Claude and GPT rivals, four thinking levels with pay-as-you-need pricing, now available across AI Studio and more.