AI & TechnologyTrend Dispatch (Pincer)
A lobster film editor at a flatbed editing table splices filmstrips of split roasted sweet potatoes while watching a brass hand-tally counter, a cloth-covered shop till behind her.

JD Researchers' Ad Agent Learned Its Edit From Clicked Videos, and Its Live Test Counted Only Clicks

CutAgent, from JD.com researchers, plans product ads in the beat order of ads that drew clicks, cuts them from each product's earlier footage and was trained against a click predictor. Its one-week live test counted clicks alone, and none of its six quality scores penalises a claim the listing does not support; one example ad promises a freshness that the customer feedback beside it says may vary.

Parallax Pincer

The sweet-potato ad opens on narration about a hot, freshly broken-open roasted potato, then promises tubers dug in season and shipped fresh, freshness maxed out (新鲜度拉满). The customer feedback printed beside it in the paper says freshness, size consistency and actual weight “may vary”. The agent learned its sequences from ads that had already drawn clicks, and its live test counted clicks alone. What those clicks bought, the paper never says.

CutAgent, posted on 17 November by researchers from JD.com and three universities, uses two agents. GeneAgent reads a product’s listing, review summary and photos and writes a “creative gene”: an ordered run of beats from a 27-term vocabulary, plus production rules. Other models reverse-extracted each training gene from a product’s earlier ads, reading every shot “with voice-over as the primary cue and visuals as auxiliary evidence”. Fine-tuning used only the top fifth of about 70,000 such products by click rate.

Reinforcement learning (GRPO) then rewarded drafts mainly by the click tier a frozen LightGBM classifier predicted, most for a category’s top 10%. The classifier’s roughly 250 features include beats, keywords, text, category and price; the paper lists no visual one. It scored 76.9% on held-out data, against 60% for always guessing the majority class. A diversity term discourages one product’s eight drafts from repeating each other. The vocabulary keeps hard-sell slots: “Price Shock” (“Only $9.9”), countdowns, an arrow with “Click the link.”

The sweet-potato ad out-clicked a rival’s cut, 2.81% to 1.27%, a gain the authors suggest “stems from narrative ordering rather than from shot-level aesthetics”. To an editor, the order is the look. VideoAgent turns the gene into shots, minding stored fixes (“avoid repeated close-ups in the opening”), and generates no footage. Each shot is picked from about 90 to 120 clips cut from the product’s earlier ads, and the paper never says who filmed them.

Filmed clips spare CutAgent most of the product errors that our AdSpark piece and Anoki’s reviewers found in generated video. A reflection loop then re-edits each video against six scores, all judged by an LLM, Qwen3.5-397B. The authors call one of them, visual-script consistency, “best read as a relative signal between methods rather than an absolute measurement”.

That score asks whether the picture “conveys the expressed meaning” of the narration. Material quality docks blur, noise, bad exposure and repetition, and asks for relevance to the target product. Selling-point coverage counts the listing’s selling points a video covers, which tests whether a point is mentioned, not whether it is true. None is described as comparing frames with the product photos.

Trust Building (TB) asks whether the video “makes the product appear reliable and worth purchasing”. Half its 100 points go to realistic usage scenes, half to evidence “such as certifications, reviews, sales statistics, or real-person demonstrations”. In the sweet-potato ad, the two beats labelled Trust Building carry the freshness promise and a closing call to click the link and order (快点击链接下单). Table 3 lists the same two codes as Gift or Bundle Value and User Testimonial.

Trust, in these tables, is a look.

The week-long live test ran on “a large e-commerce platform”, unnamed, with randomly allocated traffic: 43 categories, ten products each, over 10 million impressions. The untuned base model, Qwen3-VL-8B, drew 1.12%; fine-tuning lifted it to 1.76%, up 0.64 points or 57%. The click reward added 0.29 points (16%); reflection and memory took it to 2.21%. That is 1.09 points above the base, nearly double, and 0.23 points (12%) above the best rival, with no confidence intervals.

People liked the result too: 30 participants made 1,600 forced choices on 200 products and picked CutAgent’s video 60.4% of the time. Yet clicks and judgement part ways. AutoCut, a rival trained on videos picked by click and like rates, led the rivals on clicks at 1.98% but fared worst on TB and in the human vote.

The paper calls its planning “conversion-oriented” and never measures a conversion. AgenticGen, a comparable loop tested in TikTok’s advertising system, reported conversion rate as well as clicks. CutAgent cannot say whether a better hook sold more sweet potatoes or only drew more clicks.

Apparel is one of the 43 categories, ten products each in the live test, yet the paper shows none of their ads. Its cleaning step removes clips “containing on-camera humans”, and the authors concede that retrieval “pays in coverage”. A dress sells on drape, and the paper never shows what is left to drape it on.

Home shopping tunes its pictures to a live signal. Diane von Furstenberg’s two-hour QVC appearance took $1.2 million for her silk separates, Adweek reported in March 1993. In 2017 Philadelphia Magazine watched a QVC line producer see calls spike, with viewers keen on a blanket’s size. Over the headset went a request to wrap it around host and guest “to show how big it is”. A click carries no question about size, and CutAgent’s live test counted nothing but clicks.