One Model, Two Markets: Bid-Aware Generative Recommendation

Yanchen Jiang; Zhe Feng; Christopher P. Mah; Aranyak Mehta; Di Wang

cs.IRMar 23, 2026

One Model, Two Markets: Bid-Aware Generative Recommendation

Authors: Yanchen Jiang, Zhe Feng, Christopher P. Mah, Aranyak Mehta, Di Wang

Organizations: Harvard University · Work done during an internship at Google Research · Google Research

Abstract

Generative Recommender Systems using semantic ids, such as TIGER (Rajput et al., 2023), have emerged as a widely adopted competitive paradigm in sequential recommendation. However, existing architectures are designed solely for semantic retrieval and do not address concerns such as monetization via ad revenue and incorporation of bids for commercial retrieval. We propose GEM-Rec, a unified framework that integrates commercial relevance and monetization objectives directly into the generative sequence. We introduce control tokens to decouple the decision of whether to show an ad from which item to show. This allows the model to learn valid placement patterns directly from interaction logs, which inherently reflect past successful ad placements. Complementing this, we devise a Bid-Aware Decoding mechanism that handles real-time pricing, injecting bids directly into the inference process to steer the generation toward high-value items. We prove that this approach guarantees allocation monotonicity, ensuring that higher bids weakly increase an ad's likelihood of being shown without requiring model retraining. Experiments demonstrate that GEM-Rec allows platforms to dynamically optimize for semantic relevance and platform revenue.

Figures & tables

Figure 1: GEM-Rec Unified Architecture. Left: training on unified organic and ad logs and sampling the slot type. Right: bid-aware decoding for ads; organic decoding stays relevance-based.

Panel of Figure 2 — Figure 2: Macro-Dynamics of GEM-Rec. on the Steam dataset (a) The trade-off between Platform Utility (Revenue) and Policy Fit (NDCG with respect to the training data). We observe a frontier where significant platform utility can be generated before the model deviates significantly from the marketplace policy. (b) The parameter $\lambda$ provides monotonic control over Ad Rate.

		Economic Utility		Total Metrics		Metrics for Organic Gen.
Dataset	Operating Mode	Ad Rate	Revenue	NDCG@10	Recall@10	O.NDCG@10	O.Recall@10
Steam (Train Ad %: 3.49)	TIGER (Baseline)	0.0%	-	0.1442 †	0.1818	0.1487	0.1875
	GEM-Rec ( $\lambda=0.0$ )	2.5%	535	0.1411	0.1782	0.1468	0.1857
	GEM-Rec ( $\lambda=1.0$ )	4.7%	1,173	0.1381	0.1742	0.1467	0.1853
[1pt/2pt] Sports (Train Ad %: 6.85)	TIGER (Baseline)	0.0%	-	0.0176 †	0.0308	0.0187	0.0327
	GEM-Rec ( $\lambda=0.0$ )	4.7%	804	0.0173	0.0335	0.0187	0.0364

Table 1: Main Results: Marketplace Performance. Comparisons of GEM-Rec against the baseline (TIGER). We report separate columns for Economic Utility, Total Metrics, and Metrics for Conditional Generation to isolate the impact of bid modulation. Train Ad % refers to the realized frequency of ads in the training sequence. See full table in Appendix G , Table 3 .

Panel of Figure 3 — Figure 3: Micro-Analysis of Recommendation Quality on the Steam dataset. (a) A validation of our central hypothesis: while NDCG drops as we force ads (Red line), the Conditional Organic NDCG (Green line) remains relatively flat. This shows that increasing $\lambda$ changes when ads are shown, but does not degrade the model’s understanding of user intent for organic slots. (b) The cost of monetization: higher welfare requires displaying ads with lower semantic similarity to the user’s organic target.

Setting	Ad Rate	High-Value Share (% of Ads)	Revenue
Baseline ( $\lambda=0.0$ )	2.4%	21.8%	$1.0\times$
GEM-Rec ( $\lambda=0.5$ )	7.1%	81.5%	$9.0\times$
GEM-Rec ( $\lambda=1.0$ )	18.0%	97.4%	$28.2\times$
GEM-Rec ( $\lambda=2.0$ )	62.4%	99.9%	$104.6\times$

Table 2: Response to Market Volatility (Steam). We analyze the composition of generated ads before and after a price shock. At

\lambda=0.5

, the system aggressively targets the high-bidders. Note that while the total Ad Rate remains low (7.1%), the High-Value Share among ads dominates, proving the model is substituting low-value inventory for high-value inventory.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 4 — Figure 4: Full Dynamics on Beauty Dataset. Macro-dynamics (top row) and Micro-analysis (bottom row) showing consistent behavior with the Steam dataset.

Panel of Figure 5 — Figure 5: Full Dynamics on Sports Dataset.

Panel of Figure 6 — Figure 6: Full Dynamics on Toys Dataset.

		Train	Economic Utility		Total Metrics		Metrics for Organic Gen.
Dataset	Method	Ad %	Ad Rate	Revenue	NDCG@10	Recall@10	O.NDCG@10	O.Recall@10
Beauty	TIGER (Baseline)	4.23%	0.0%	0.0	0.0282 †	0.0529	0.0293	0.0550
	GEM-Rec ( $\lambda=0.0$ )		3.1%	345.1	0.0301	0.0569	0.0318	0.0602
	GEM-Rec ( $\lambda=0.5$ )		4.2%	489.3	0.0296	0.0560	0.0317	0.0603
	GEM-Rec ( $\lambda=1.0$ )		6.0%	726.6	0.0295	0.0559	0.0320	0.0606
	GEM-Rec ( $\lambda=2.0$ )		10.7%	1,386.8	0.0282	0.0537	0.0319	0.0610

Appendix

Table 3: Full Marketplace Performance Results. We compare GEM-Rec against the baseline (TIGER) across all datasets and

\lambda

settings. The Train Ad % column is provided for reference.

Dataset	Valid Rate
Beauty	100.0%
Sports	100.0%
Toys	100.0%
Steam	100.0%

Appendix

Table 4: Generative Validity. The percentage of generated ad sequences that decode to a valid, existing item in the inventory.

Setting	Ad Rate	High-Value Share	Revenue Uplift
Baseline ( $\lambda=0$ )	2.38%	21.8%	$1.0\times$
GEM-Rec ( $\lambda=0.5$ )	7.01%	81.5%	$9.0\times$
GEM-Rec ( $\lambda=1.0$ )	18.03%	97.4%	$28.2\times$
GEM-Rec ( $\lambda=2.0$ )	62.44%	99.9%	$104.6\times$

Appendix

Table 5: Bid Shock Response (Steam). The baseline (

\lambda=0

) relies on historical priors and misses the new opportunity. GEM-Rec (

\lambda\geq 0.5

) dynamically pivots to capture the high-value inventory.

Setting	Ad Rate	High-Value Share	Revenue Uplift
Baseline ( $\lambda=0.0$ )	3.49%	3.2%	$1.0\times$
GEM-Rec ( $\lambda=0.5$ )	10.31%	49.3%	$14.3\times$
GEM-Rec ( $\lambda=1.0$ )	24.45%	81.7%	$55.8\times$
GEM-Rec ( $\lambda=2.0$ )	68.66%	97.2%	$202.4\times$

Appendix

Table 6: Bid Shock Response (Toys). At

\lambda=0

, the High-Value Share is extremely low (3.2%), indicating the model is effectively blind to the new bids. GEM-Rec corrects this rapidly, reaching 97% share by

\lambda=2.0

Setting	Ad Rate	High-Value Share	Revenue Uplift
Baseline ( $\lambda=0.0$ )	3.12%	1.6%	$1.0\times$
GEM-Rec ( $\lambda=0.5$ )	8.76%	43.0%	$14.7\times$
GEM-Rec ( $\lambda=1.0$ )	20.51%	77.6%	$57.1\times$
GEM-Rec ( $\lambda=2.0$ )	63.65%	98.5%	$240.6\times$

Appendix

Table 7: Bid Shock Response (Beauty). The model demonstrates consistent adaptability. The High-Value share jumps from a negligible 1.6% to 43% with only a minor increase in Ad Rate (from 3.1% to 8.8%).

Setting	Ad Rate	High-Value Share	Revenue Uplift
Baseline ( $\lambda=0.0$ )	4.87%	4.7%	$1.0\times$
GEM-Rec ( $\lambda=0.5$ )	12.56%	72.9%	$14.6\times$
GEM-Rec ( $\lambda=1.0$ )	30.41%	99.2%	$50.1\times$
GEM-Rec ( $\lambda=2.0$ )	74.71%	100.0%	$136.2\times$

Appendix

Table 8: Bid Shock Response (Sports). This dataset exhibits the most aggressive adaptation. At

\lambda=1.0

, nearly every ad shown (99.2%) is already from the high-value subset.

Panel of Figure 7 — Figure 7: Dynamics on Steam (High Ad Rate). The model correctly learns the higher base ad rate ( $\sim$ 10%) at $\lambda=0$ and allows strictly monotonic boosting beyond this point.

Panel of Figure 8 — Figure 8: Dynamics on Beauty (High Ad Rate).

Panel of Figure 9 — Figure 9: Dynamics on Sports (High Ad Rate). Note here that the drop of the green dotted curve at $\lambda=10$ for figure c) here corresponds with a nearly 100% ad rate (see figure b) leaving very few organic generation, therefore it reflects a high variance and does not contradict our main conclusion and observations.

Panel of Figure 10 — Figure 10: Dynamics on Toys (High Ad Rate).

		Train	Economic Utility		Total Metrics		Metrics for Organic Gen.
Dataset	Method	Ad %	Ad Rate	Revenue	NDCG@10	Recall@10	O.NDCG@10	O.Recall@10
Beauty	TIGER (Baseline)	11.99%	0.0%	0.0	0.0245 †	0.0458	0.0277	0.0518
	GEM-Rec ( $\lambda=0.0$ )		10.5%	1,137.1	0.0314	0.0565	0.0322	0.0589
	GEM-Rec ( $\lambda=0.5$ )		14.3%	1,653.0	0.0302	0.0546	0.0317	0.0581
	GEM-Rec ( $\lambda=1.0$ )		18.7%	2,285.2	0.0301	0.0539	0.0315	0.0578
	GEM-Rec ( $\lambda=2.0$ )		30.0%	3,953.5	0.0286	0.0517	0.0317	0.0587

Appendix

Table 9: Full Marketplace Performance Results (Strict Last-Item Protocol). We compare GEM-Rec against the baseline (TIGER) across all datasets and

\lambda

settings under the new evaluation setting.

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 502 papers

Generative Recommendation for Large-Scale Advertising
Feb 26, 2026Ben Xue, Dan Liu, Lixiang Wang +27RL for Recommendation Recommender Systems
Kuaishou Technology, Beijing, China
Token-Level Advertising
Aug 27, 2026Hanbing Liu, Bowei Zhang, Changyuan Yu +2Online Advertising Mechanism Design
Renmin University of China Beijing, China · Beijing, China · Stanford University Stanford, CA, USA
JD-BP: A Joint-Decision Generative Framework for Auto-Bidding and Pricing
Apr 7, 2026Linghui Meng, Chun Gan, Shengsheng Niu +8Automated Bidding Real-Time Bidding
JD.com China · The Chinese University of Hong Kong China

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

Generative Recommendation for Large-Scale Advertising

Ben Xue, Dan Liu, Lixiang Wang, Mingjie Sun+26

Kuaishou Technology, Beijing, China

RL for Recommendation Recommender Systems Generative Recommendation Online Advertising

Open paper AskPDF arXiv

Generative recommendation has recently attracted widespread attention in industry due to its potential for scaling and stronger model capacity. However, deploying real-time generative recommendation in large-scale advertising requires designs beyond large-language-model (LLM)-style training and serving recipes. We present a production-oriented generative recommender co-designed across architecture, learning, and serving, named GR4AD (Generative Recommendation for ADdvertising). As for tokenization, GR4AD proposes UA-SID (Unified Advertisement Semantic ID) to capture complicated business information. Furthermore, GR4AD introduces LazyAR, a lazy autoregressive decoder that relaxes layer-wise dependencies for short, multi-candidate generation, preserving effectiveness while reducing inference cost, which facilitates scaling under fixed serving budgets. To align optimization with business value, GR4AD employs VSL (Value-Aware Supervised Learning) and proposes RSPO (Ranking-Guided Softmax Preference Optimization), a ranking-aware, list-wise reinforcement learning algorithm that optimizes value-based rewards under list-level metrics for continual online updates. For online inference, we further propose dynamic beam serving, which adapts beam width across generation levels and online load to control compute. Large-scale online A/B tests show up to 4.2% ad revenue improvement over an existing DLRM-based stack, with consistent gains from both model scaling and inference-time scaling. GR4AD has been fully deployed in Kuaishou advertising system with over 400 million users and achieves high-throughput real-time serving.

Abstract

Recommender Systems

Explore similar work

Generative Recommendation for Large-Scale Advertising

Token-Level Advertising

JD-BP: A Joint-Decision Generative Framework for Auto-Bidding and Pricing