G-CAST: Governed Continual Self-Training for E-Commerce Agents under Evolving Demands
E-commerce agents do not operate against a permanently fixed target. Customers ask for new kinds of products and procurement workflows, while the agent itself gains tools, skills, memory, and operating procedures. The authors of G-CAST argue that this creates a continual adaptation problem involving three linked mismatches: the model may not know how to use new capabilities, old training data may not cover them, and the reward function may no longer recognize correct behavior.
Why ordinary retraining is not enough
A model trained under an earlier capability contract can deteriorate as the surrounding harness expands. The paper defines that contract as the information available to the agent, the actions it can take, and the interaction protocols it must follow. Adding a tool does not guarantee that the model will invoke it at the right time or integrate its output correctly.
Periodic manual retraining is also poorly matched to production change. It can be slow and labor-intensive, and optimizing only for new tasks risks damaging behaviors that existing users already depend on. G-CAST therefore treats an accepted model checkpoint, rather than an isolated training run, as the unit of adaptation. A candidate must improve on emerging demand while retaining earlier capabilities.
How G-CAST works
G-CAST is a governed loop with two connected parts. The main loop collects evidence from daily production queries, generates multiple rollouts, selects training examples, adapts the model, and evaluates a candidate against both new and retained tasks. The reward-evolution loop is activated when evidence suggests that the current reward is no longer reliable under the active capability contract.
For each query, the system generates eight rollouts from the accepted checkpoint. It routes queries according to their reward outcomes. Queries whose rollouts all fail can enter supervised fine-tuning when a verified target exists. Mixed-success queries enter reinforcement learning because they provide useful variation. Queries whose rollouts all succeed are held for regression monitoring rather than treated as obvious training material.
Within the supervised and reinforcement-learning pools, G-CAST prioritizes examples using token-level entropy. The important design choice is that reward determines a query’s training role, while entropy ranks examples within that role. In the authors’ formulation, uncertainty is not treated as a substitute for correctness.
Reward evolution addresses a different failure mode. When the contract changes or audits reveal suspicious scoring, training pauses. The proposed executable reward update must pass targeted validation, regression checks, offline rescoring, and trajectory replay before it can control training. Candidates are trained from an immutable accepted parent and promoted only after cumulative evaluation.
Evidence from a deployed procurement agent
The authors evaluated G-CAST in Yuancai, a B2B procurement agent on Alibaba’s 1688 platform. They studied four successive demand and capability periods, expanding from four tools in the first period to 23 in the fourth. Each period used a frozen rollout pool and a separate evaluation slice, with cumulative evaluation growing as new periods were introduced.
On the unchanged original checkpoint, cumulative base reward fell from 0.9065 to 0.6354 across the four-period evaluation, while reward on the newly introduced fourth period fell to 0.1877. This supports the paper’s central premise: harness expansion alone does not make an older checkpoint effective on new demand.
G-CAST was compared with AutoTrainess under matched cumulative evaluation. In the third iteration, G-CAST reached a cumulative reward of 0.9288 after reinforcement learning, compared with 0.9210 for AutoTrainess. In the fourth, it reached 0.8078 versus 0.7911. The paper also reports a better retention-adaptation balance in the fourth iteration: G-CAST raised retained reward from 0.8712 to 0.8833 while new-period reward rose from 0.3983 to 0.5813. AutoTrainess improved the new period but retained reward declined from 0.8907 to 0.8739.
The authors attribute much of the later advantage to behaviors required by the expanded contract. In the fourth iteration, G-CAST achieved 0.7625 skill-selection accuracy versus 0.6875 for AutoTrainess, and skill adherence of 0.4503 versus 0.4219.
What the ablations show
The routing ablation separates reward routing from entropy prioritization. On the four-period cumulative evaluation, post-reinforcement-learning reward was 0.7930 with the full Reward plus Entropy strategy, 0.7721 with Reward plus Random selection, and 0.7334 with Random selection alone. The result suggests that routing supplies a useful learning signal and that entropy-based ranking adds value within routed pools.
The reward-evolution ablation points to governance as more than a cosmetic safeguard. With the same training data and optimization settings, the gated reward design reached 0.9288 cumulative reward, compared with 0.8926 for an ungated one-time update and 0.8810 for a static reward. On the skill-intensive third period, the gated approach reached 0.9538, while static and ungated rewards reached 0.8801 and 0.8812.
Audits also found high false-positive rates among sampled high-reward trajectories: 84% for search-result misuse and 82% for hallucinated answer entities. These cases motivated a grounding penalty for unsupported identifiers, URLs, prices, or phone numbers. The paper reports that validation gates rejected three candidate checkpoints and two reward versions during deployment.
Limitations and practical implications
The study is based on one longitudinal production application and one primary model family. Its demand patterns, reward design, and governance process may not transfer directly to other agents. The reported improvements also come from the authors’ particular evaluation protocol and contract definitions, so they should not be read as universal evidence that every continual-training system will benefit equally.
For practitioners, the broader lesson is architectural. Continual adaptation should version the model, capability contract, reward, training evidence, and evaluation history together. Production traces should be routed according to what they can teach, not simply pooled. New reward rules should be tested against known-good and known-bad behavior before they influence training. Finally, promotion gates should measure both emerging-task performance and regression on retained capabilities.
G-CAST does not eliminate the need for human judgment, but it attempts to make that judgment more targeted and auditable. The authors report reducing reward-audit effort from about two person-days to 30 minutes per round and shortening demand-to-acceptance time from about one week to two days. Those operational results are promising, while the single-application scope leaves open how well the approach will generalize beyond this e-commerce setting.
Sources
AI-assisted research summary. Check the original paper for complete methods, results, and limitations.