NextAds

Envision Next-Generation Personalized Video Advertising

Yiyan Xu1, Ruoxuan Xia1, Wuqiang Zheng1, Fengbin Zhu2, Wenjie Wang1, Fuli Feng1
1 University of Science and Technology of China, Hefei, China 2 National University of Singapore, Singapore
Closed-loop Creative Refinement

Two Feedback Pathways

Before deployment

Pre-deployment Verification

Before a creative is served, Verifier identifies alignment, execution, or compliance failures and routes diagnostic evidence to the smallest responsible stage for targeted repair.

After deployment

Iterative Refinement Across Impressions

After deployment, Reflector converts online interaction signals into update directives that refine the creative shown in subsequent impressions.

Pre-deployment Verification

Verifier findings are routed to the smallest responsible stage before a creative is served.

Matcher repair

Coca-Cola · Replace a Competing Interest Topic

Matcher selected the user’s medium CHAGEE beverage interest instead of the stronger crochet and handmade interest, creating a competing-brand bridge for Coca-Cola.

ProblemThe video uses tea-brand-like props and beverage cues that conflict with Coca-Cola and miss the user’s strongest preference.
Send toMatcher reruns user–product matching; Director rewrites the affected scenes; Producer regenerates the conflicting segment.
ChangeUse crochet and handmade as the personalization anchor, remove third-party beverage cues, and place Coca-Cola beside yarn, crochet flowers, and handmade props.
BeforeCHAGEE-related beverage topic
AfterCrochet / handmade preference
Product–preference bridge reconstruction

Airbnb · Reframe a Tea Interest Around the Stay

The user’s tea interest is used too literally, so the creative looks like a tea-shop promotion instead of an Airbnb advertisement.

ProblemTea cups, packaging, postcards, and cafe cues dominate, while discovering, choosing, and booking an Airbnb stay are not clearly shown.
Send toMatcher rebuilds the product–preference bridge; Director rewrites the story around Airbnb stay discovery; Producer removes third-party beverage cues.
ChangeUse tea only as local-culture ambience around browsing a stay, receiving host guidance, arriving at the room, and planning the trip.
BeforeTea-shop-like personalization
AfterAirbnb stay discovery
Localized video editing

Coca-Cola · Repair Two Faulty Segments

Verifier identifies two timestamped execution failures. Producer edits only those segments, stitches them back into the video, and keeps the original audio.

ProblemAt 2.0–4.8s, idol/TXT-like cards and event text imply an unauthorized association; at 5.0–7.4s, the person appears to drink while the bottle cap is still attached.
Send toProducer receives two localized editing instructions for the affected timestamps.
ChangeReplace the recognizable K-pop assets with generic concert props and repair the drinking action so the cap is visibly removed before the sip.
BeforeAssociation and product-use errors
AfterTwo localized repairs
User-profile regeneration

Airbnb · Rebuild the Profile from Full History

The original profile is dominated by medical-aesthetic interests that do not provide a safe, natural connection to Airbnb.

ProblemAcne-scar repair, microneedling, laser, and treatment-recovery topics should not be forced into a lodging advertisement.
Send toProfile Builder searches all 45 recent clicked notes and rebuilds the profile; Matcher then evaluates the new product-conditioned preference tiers.
ChangeSelect the supported “evidence-driven and anti-misleading decision making” preference and connect it to real listing photos, reviews, host notes, location details, and transparent booking information.
BeforeNo safe Airbnb preference bridge
AfterEvidence-driven booking preference

Iterative Refinement Across Impressions

Reflector converts online interaction signals into update directives for subsequent impressions.

Negative feedback

iPhone 17 · Opening-hook Refinement

A simulated user skips at 2.8s. Reflector treats the slow opening as a possible cause and limits the revision to the first three seconds.

FeedbackSkip at 2.8s; no like, click, share, or comment.
InterpretThe opening may be weak or too slow because the user leaves before the product-use demonstration; confidence: 0.72.
Send toDirector rewrites the 0–3s hook; Producer edits only that segment.
ChangeShow an immediate iPhone 17 use moment within the first two seconds. Keep the later 3–15s sequence and final end card unchanged as a minimal-edit constraint.
BeforeSlow opening setup
AfterImmediate product-use hook

Citation

arXiv:2603.02137 · DOI: 10.48550/arXiv.2603.02137Paper ↗
@article{xu2026nextads,
  title={Envision Next-Generation Personalized Video Advertising},
  author={Xu, Yiyan and Xia, Ruoxuan and Zheng, Wuqiang and Zhu, Fengbin and Wang, Wenjie and Feng, Fuli},
  journal={arXiv preprint arXiv:2603.02137},
  year={2026},
  doi={10.48550/arXiv.2603.02137}
}