Exploratory creative pre-ranking: Hierarchical empirical-Bayes Thompson sampling for real-time creative optimization

Jiale Han; Ajith Moparthi; Dan Xie; Xiaowu Dai

cs.LGOct 26, 2026

Exploratory creative pre-ranking: Hierarchical empirical-Bayes Thompson sampling for real-time creative optimization

Authors: Jiale Han, Ajith Moparthi, Dan Xie, Xiaowu Dai

Organizations: Department of Statistics and Data Science, UCLA · Roblox · Departments of Statistics and Data Science and of Biostatistics, UCLA

Abstract

Dynamic creative optimization decides which creative of an ad to render for each request. Recent work places that choice before ad ranking, but relies on deterministic point predictors with no calibrated uncertainty. AI-generated content has made creatives nearly free, while feedback per creative remains sparse and selection faces a tight latency budget. We introduce a creative pre-ranker that scores every recalled creative and shortlists each campaign's pool before the per-user ranker runs. Creatives sit in a hierarchy of audience segment, advertised item, campaign and creative, so a cold creative falls back on what its parent levels have earned, and how much it borrows from its parents is set in closed form from historical logs, with one parameter per level and nothing tuned. To stay within the latency budget, the pre-ranker predicts for an audience segment rather than an individual user, from per-attribute counters. It selects by a single Thompson draw per creative, balancing exploration and exploitation so that a new creative still has a chance to be learned. Replaying Roblox Sponsored-ads traffic over campaigns with competing creatives, the shortlist lifts predicted play-through rate by 8.14% against 2.80% for an equally sized random shortlist, forwards more than 40% fewer creatives than the full pool, and adds only 0.2% to serving time. After an A/B test on live traffic, the method now serves in production at Roblox.

Figures & tables

Panel of Figure 1 — Figure 1: Two different creatives of a single advertised item.

Panel of Figure 2 — Figure 2: The creative pre-ranker in context. It scores every recalled creative with one draw from its hierarchical posterior, assembled offline and refreshed online from single-attribute counters, and shortlists the top $E_{g}=\max(1,\lfloor\rho|\mathcal{C}_{g}|\rfloor)$ of each campaign (inset) up to a global budget $B$ filled by round robin. Only that shortlist reaches ad ranking, which scores it with the production creative-selection model and keeps at most one creative per advertised item. The cap $B$ controls its per-candidate cost.

\quad\quad\quad\quad\ \ \begin{aligned} \tilde{\theta}_{c,s}\sim\mathrm{Beta}\big(&\kappa_{c}\mu+n_{\mathrm{eff}}\hat{\mu}_{c,s},\kappa_{c}(1-\mu)+n_{\mathrm{eff}}(1-\hat{\mu}_{c,s})\big).\end{aligned}

Algorithm 1 Exploratory Creative Pre-Ranking by Hierarchical Empirical-Bayes Thompson Sampling.

Method	Forwarded	Lift ( $95\%$ CI)	Gap
Uniform	$\approx\!L$	—	$0.0\%$
RR-Random	$600.0$	$+2.80\ [2.71,\,2.90]$	$15.3\%$
TS-One	$\approx\!L$	$+6.09\ [5.62,\,6.58]$	$33.2\%$
TS-RR	$600.0$	$+8.14\ [7.64,\,8.65]$	$44.3\%$
Mean-RR	$600.0$	$+9.51\ [8.87,\,10.18]$	$51.8\%$
Oracle	$>1{,}000$	$+18.36\ [17.78,\,18.97]$	$100\%$

Table 1: Predicted play-through rate on contested pools,

\rho=0.8

B=600

. Lifts are against Uniform with

95\%

request-clustered bootstrap intervals; lift and gap follow ( 24 ). Forwarded is the mean creatives per request reaching the ranker.

		Pre-ranker ( $\mu$ s)		Ranker	Total
Method	Fwd.	Score	Select	(ms)	(ms)
Uniform	$\approx\!L$	$0.1$	$14.8$	$33.0$	$33.0$
RR-Random	$600.0$	$6.6$	$2{,}109.9$	$39.4$	$41.5$
TS-One	$\approx\!L$	$78.0$	$2{,}071.0$	$33.0$	$35.1$
TS-RR	$600.0$	$79.2$	$2{,}122.5$	$39.4$	$41.6$
Oracle	$>1{,}000$	$0.1$	$3.0$	$83.1$	$83.1$

Table 2: Per-request computation time. Score and select are the pre-ranker’s two timed phases, and total adds the ranker the cap protects. Oracle ’s ranker entry is extrapolated past the largest shortlist measured.

$\rho$	$B$	Fwd. ( $\times L$ )	RR-Random	TS-RR	Gap
$0.2$	$600$	$1.00$	$+0.02\%$	$+3.34\%$	$32.5\%$
$0.4$	$600$	$1.03$	$+0.26\%$	$+3.52\%$	$34.2\%$
$0.6$	$600$	$1.15$	$+1.14\%$	$+4.17\%$	$40.5\%$
$0.8$	$500$	$1.01$	$+0.04\%$	$+3.39\%$	$33.0\%$
$0.8$	$600$	$1.21$	$+1.62\%$	$+4.52\%$	$44.0\%$
$0.8$	$800$	$1.57$	$+4.09\%$	$+6.29\%$	$61.2\%$

Table 3: Quota fraction

\rho

against latency cap

B

under another deployed ranking model, with

50

draws per candidate. Lifts are against Uniform and gap is TS-RR ’s, as in ( 24 ). Forwarded is in multiples of the one-creative-per-campaign shortlist.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Recalled pools that are contested ( $\|\mathcal{C}_{g}\|\geq 2$ )	$99.1\%$
Campaigns per request, mean	just under $L$
Creatives per request, mean	$>1{,}000$
Creatives per campaign, mean	$2.637$
Creatives per campaign, max	$20$
Cold-start candidates ( $n_{\mathrm{eff}}=0$ )	$16.9\%$

Appendix

Table 4: The replay sample, as shares and per-request or per-campaign rates rather than totals. A pool is a (request, campaign) pair, contested when the campaign has at least two recalled creatives.

Method	Forwarded	Lift ( $95\%$ CI)	Gap
Uniform	$\approx\!L$	—	$0.0\%$
RR-Random	$600.0$	$+1.62\ [1.51,\,1.72]$	$15.7\%$
TS-One	$\approx\!L$	$+3.35\ [3.01,\,3.69]$	$32.5\%$
TS-RR	$600.0$	$+4.54\ [4.16,\,4.91]$	$44.0\%$
Mean-RR	$600.0$	$+5.10\ [4.68,\,5.52]$	$49.5\%$
Oracle	$>1{,}000$	$+10.30\ [9.68,\,10.92]$	$100\%$

Appendix

Table 5: Robustness under L1. The setup matches Table 1 .

Learning to Rank

Learning to rank develops models that order items by relevance or utility for search, recommendations, and other decisions.

See all 273 papers

Offline-to-Online Creative Optimization with Generative Models and Adaptive Testing
Jul 26, 2026Kevin Lee, Benjamin Letham, Zhiyuan Jerry Lin +5Online Advertising Generative Modeling
University of Michigan · USA · Meta
PinDCO: Whole-Page Aware Dynamic Creative Optimization at Scale
Sep 14, 2026Yu Hao, Yuchun Li, Peimeng Sui +5Learning to Rank Online Advertising
Pinterest Inc.
Too Late to Prune: Generative AI, the Screen, and Team Creative Decisions
Oct 10, 2026Difang Huang, Shumin Yu, Fadong Chen +1Human-AI Decision Making Human-AI Co-Creation
Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China · School of Management, Zhejiang University, Hangzhou, China

Jul 26, 2026J/K move · Enter open · S save

Offline-to-Online Creative Optimization with Generative Models and Adaptive Testing

Kevin Lee, Benjamin Letham, Zhiyuan Jerry Lin, Elodie Samson+4

University of Michigan · USA · Meta

Online Advertising Generative Modeling A/b Testing

Open paper AskPDF arXiv

Ad creative optimization is increasingly constrained by evaluation rather than generation. Generative models can produce many plausible creatives, but reliable evaluation requires online experiments, in which only a limited slate can be tested. We study how to use data from historical A/B tests to generate and select the candidates in that slate. We developed and deployed a performance-driven offline-to-online workflow that guides creative generation with a predictive model as an inference-time critic. In the offline phase, we use a predictive model trained on historical experiments to rank and refine variants created by a generative model. A final test slate is then deployed in an online adaptive experiment. In a 50-arm field experiment, we found that the best creative generated with this method yielded 45.1% higher engagement than the best human-authored creative. Two additional experiments showed the same upper-tail pattern, with lifts of 46.7% and 36.2%. We found that despite the predictive model being too noisy to directly identify the best creative offline, it effectively guides the generative model toward creating strong candidates that can be efficiently evaluated in an adaptive experiment. The results suggest a design principle for creative optimization with generative models: use predictive models to guide generation of a slate to test, judge the slate by whether it contains high-performing candidates at a feasible test size, and use adaptive experiments to select among candidates while limiting traffic lost to weak arms.

Abstract

Learning to Rank

Explore similar work

Offline-to-Online Creative Optimization with Generative Models and Adaptive Testing

PinDCO: Whole-Page Aware Dynamic Creative Optimization at Scale

Too Late to Prune: Generative AI, the Screen, and Team Creative Decisions