
Alibaba Retrains 1688's Factory-Finding Agent on Buyers' Queries, and No Test of It Counts an Order
An Alibaba paper says G-CAST, now deployed in 1688's Yuancai buying workbench, retrains the agent on buyers' daily queries under a reward that is itself revised, with the human reward audit cut to 30 minutes a round. None of the measures it defines is a buyer's order, sample or dispute, and the site's one notice of a model change names no model.
Neritus Vale
An Alibaba paper says the agent that shortlists factories on 1688, Alibaba’s wholesale site, is retrained on buyers’ daily queries and graded by a rule that is itself revised. The input box of its workbench, Yuancai (源采), prompts visitors with a request: reputable factories near Zhejiang that make custom summer T-shirts in small runs and send free samples. None of the measures the paper defines is an order, an approved sample or a dispute, so treat a Yuancai shortlist like a supplier’s quote: dated, filed and checked.
G-CAST, submitted to arXiv on 13 November by four Alibaba Group authors, “is now deployed in Yuancai”, the paper says. It describes Yuancai as an agent product for business buyers, with general-assistant, quick-search and knowledge-assisted modes.
Each G-CAST round replays selected production queries through the accepted model, eight answers apiece. A query whose eight answers all score below the round’s reference score goes to supervised fine-tuning, if a verified answer exists. Mixed results go to reinforcement learning, and queries whose answers all score above the reference go to regression monitoring. A candidate becomes the accepted model only after a “cumulative evaluation” on new and previously accepted tasks.
The scoring rule, the paper’s reward, is treated as code that can be wrong. When answers suggest it is misjudging them, training pauses while a second loop revises it and tests the new rule offline and on an audit batch. The agent and the rule that trains it change together.
Most of what the reward grades is the agent’s behaviour, not what a buyer receives. Its format term measures “compliance with the required output and interaction rules rather than task correctness”. Other terms penalise unsupported facts and empty-handed reports of a failed search, and check that the agent picked and followed the right skill. The evaluation version adds “task-outcome” components, which the paper does not define.
The paper’s test of its data strategy is scored on that rule. In Table 3, the base model, an FP8 build of Qwen3-4B-Thinking-2507, scored 0.7331 on the pooled test queries; the paper’s Table 2 puts the same base at 0.6501 on that suite. The table files the score under “Task Performance”, but it is still the reward’s score. Fine-tuning on examples picked the paper’s way raised it to 0.7800.
On a sourcing site, the same table’s “Sel. Acc.” column is easy to misread. It is skill-selection accuracy, recording whether “the required skill is correctly selected”: which procedure the agent invoked, not which supplier it chose. With the paper’s method it reached 73.25%, a measure the authors say “differs only slightly” across data strategies.
When the abstract promises new behaviour “without sacrificing previously accepted capabilities”, acceptance means a score, not a sale. The paper’s cumulative test set holds 400 production queries, frozen before training and graded by the final reward. The paper gives no breakdown by category, so it cannot say how the agent does on clothing. Yuancai’s home page nonetheless showcases a comparison of contract factories for seamless underwear (无缝内衣的代工厂对比分析, our translation).
An earlier version of the reward was caught scoring unsupported details highly. The authors’ audit of high-scoring answers, with no sample size given, found clusters with false-positive rates of 84% for search-result misuse and 82% for hallucinated entities. In each, the agent called a plausible search tool and still produced “offer identifiers, URLs, prices, or phone numbers unsupported by the returned evidence”. The grounding penalty was the response.
G-CAST also saved people’s time on the audit itself. In deployment, human reward-audit effort, the human check on the rule, fell from about two person-days to 30 minutes a round.
G-CAST cut what the paper calls “demand-to-acceptance turnaround” from about a week to two days. If acceptance means going live, the agent a buyer consults can change within days of a new kind of request.
Its gates do refuse. They turned back three candidate checkpoints and two proposed reward versions, out of a total the paper does not give. The cases, it says, are “documented in Appendix E”.
Appendix E is not in the version posted on 13 November, which ends at its references. Appendix B and the appendix tables the text cites are missing too. Readers see what the gates let through and nothing of what they stopped.
Which model answers buyers is not stated either. The experiments start from Qwen3-4B-Thinking-2507, a four-billion-parameter model published by the Qwen team. The paper never names the model serving Yuancai’s traffic, though it speaks of “the deployed model”. Its stated limits include “one longitudinal production application and one primary model family”.
Yuancai’s notice page lists seven notices; the earliest is dated 23 March, and one concerns a model. On 25 August a banner announced that “Yuancai has secretly evolved!” (源采偷偷进化啦!, our translation). A new model was live for its general assistant (新模型已上线), it said, with faster, more reliable answers. It named no model and gave no version.
The site’s code shows a model can be displayed. It holds a model switcher, labelled “current model” (当前模型), that appears only when the address contains “pre-” and a showModelSelect flag is set. The address buyers use, cai.1688.com, contains no “pre-”.
G-CAST versions the model, the reward and the evidence, and leaves the buyer to version the answer.
The strongest objection is that this is responsible maintenance. Left untouched, the original model scored 0.1877 on the base reward for the newest period’s tasks, which brought sandbox, file and export tools; the paper’s other figures use the full reward. Gated retraining spares buyers that decay, and the rejected models show the gates bite. The thesis fails if a higher reward means better factories for the buyer.
Nothing in the paper tests that link. Its grounding penalty, the term nearest a buyer’s interest, checks that an answer’s entities are supported by the request and the tools’ responses. That is not the same as a listing being true or a factory sound. As with 1688’s claim that AI starts about 40% of its demand, which never said where those requests ended, the measurement stops short of the order.
For a brand sourcing clothing on 1688, the cost is a filing job. By the banner’s account, a T-shirt request typed before 25 August and the same request typed after it could reach different general-assistant models, and the banner named neither. 1688 could print the model version beside every answer: G-CAST versions it, and the site’s code already defines a call for its name. Until it does, a shortlist’s version history is whatever the buyer writes down.