One Model, Two Markets: Bid-Aware Generative Recommendation
Generative recommender systems like TIGER excel at semantic retrieval but ignore the economic realities of monetization via sponsored content. This paper proposes GEM-Rec, a unified framework that augments semantic IDs with control tokens (<ORG>, <AD>) to factorize slot allocation from item generation, and introduces Bid-Aware Decoding to inject real-time auction bids into inference. The work bridges the gap between generative recommendation and computational advertising, offering theoretical guarantees like allocative monotonicity while allowing dynamic trade-offs between user relevance and platform revenue.
GEM-Rec presents a compelling architectural solution to an important industrial problem: unifying organic and sponsored recommendation within a single generative model. The factorization of the generative objective into Ad Satisfaction Modeling $P(f_t|H_{
The control token mechanism successfully decouples slot type decisions from content retrieval, allowing a single model to serve both markets without architectural branching. The Bid-Aware Decoding mechanism provides practical inference-time steerability via the $\lambda$ parameter, and the proofs of Allocative Monotonicity (Proposition 1) and Organic Integrity (Proposition 2) hold up under the stated assumptions. The empirical validation of "Organic Integrity" is particularly convincing: Table 1 shows Conditional Organic NDCG remains stable (e.g., 0.1857 to 0.1853 on Steam) even as $\lambda$ increases, confirming that bid modulation does not perturb organic ranking logic.
The evaluation environment is entirely synthetic: bids are drawn from Log-Normal distributions (Appendix D.1) and ad insertion follows a hand-engineered "Two-Stage Policy" with artificial frequency capping. This raises concerns about external validity, as real-world advertiser behavior involves strategic bid shading and budget constraints not captured here. The "Bid Shock" experiments (Table 2) demonstrate extreme scenarios—revenue uplifts of $104.6\times$ at $\lambda=2.0$—but this requires pushing ad rates to 62.4%, which would likely violate user experience constraints in production. Additionally, while the paper claims to learn "valid placement patterns directly from interaction logs," the training data itself is synthetic, creating a circular validation concern.
The comparison against TIGER as a "Pure Utility" baseline is methodologically sound for isolating the revenue-relevance trade-off, though it is inherently an apples-to-oranges comparison since TIGER was not designed for sponsored content. The paper correctly notes that "separate stacks" approaches (Chen et al., 2022; Yan et al., 2020) require merging at serving time, but the claim that GEM-Rec is the first to enable "direct modulation by live bids" in generative retrieval should be viewed cautiously given the concurrent nature of the field. The evidence for "100% validity rate" (Appendix G.1.4) is strong but limited to the reported datasets; it is unclear if this holds at scale or with different RQ-VAE configurations.
The authors commit to open-sourcing code after acceptance, but currently neither the GEM-Rec implementation nor the original TIGER code is publicly available. The replication relies on a third-party TIGER reproduction from Yang et al. [2024], introducing potential baseline variance. Hyperparameters are detailed in Appendix F (T5 encoder-decoder, 6 layers, $d_{model}=128$, RQ-VAE with codebook size 256), and the synthetic data generation protocol is described in Appendix D, allowing independent reconstruction of the simulation. However, the lack of real-world bid logs or production feedback loops means independent reproduction will be limited to the same synthetic environment, unable to validate the core claim of handling "real-time pricing" volatility.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.