Source-linked AI summary
Token-Level Advertising
Hanbing Liu, Bowei Zhang, Changyuan Yu, Yinyu Ye, Qi Qi
TL;DR
Generative AI creates responses dynamically, challenging advertising mechanisms built around predefined slots and motivating mechanisms that shape opportunities during generation. This paper proposes LAMA, a token-level auction with sequential incentive guarantees and near-optimal KL-regularized welfare, plus a learned online implementation. Proof-of-concept experiments report improved platform welfare and revenue while maintaining user-facing response quality.
Problem
Generative AI makes advertising opportunities emerge during response generation, limiting the fit of mechanisms designed for predefined slots.
Method
LAMA embeds advertiser influence at token level through advertiser-specific next-token policies, a latent mixture, an evolving allocation posterior, and online report reconstruction.
Results
LAMA satisfies Markov DSIC and IR, achieves near-optimal KL-regularized welfare, and improves platform welfare and revenue while maintaining user-facing response quality.
Takeaways & Limitations
The results provide initial proof-of-concept evidence that generation-native advertising can improve monetization while preserving user-facing response quality.
Takeaways & Limitations
The paper focuses on the fundamental single-winner setting, with the eventual sponsored outcome assigned to a single advertiser.
Abstract
from arXiv · showhide
Generative AI is transforming how people access information, challenging traditional advertising mechanisms built around predefined slots. Towards generation-native advertising, we propose the Latent Advertiser Mixture Auction (LAMA), a token-level advertising mechanism that embeds advertiser influence directly into the generation process. Advertisers report local continuation values that induce advertiser-specific next-token policies, from which the platform decodes through a latent mixture while updating an allocation posterior. We show that LAMA satisfies Markov DSIC and IR, and achieves near-optimal KL-regularized welfare. We further develop a learning-based implementation that reconstructs the required reports online from learned local advantages and root values. Proof-of-concept experiments on real-world commercial-search query splits show that LAMA improves platform welfare and revenue while maintaining user-facing response quality, providing initial evidence for the feasibility of generation-native advertising.
1 Introduction
Generative AI makes advertising opportunities emerge within generated responses rather than existing as predefined slots, motivating token-level mechanisms that jointly shape content, allocation, and payments. The paper proposes LAMA, proves incentive and welfare properties, develops a practical implementation, and reports proof-of-concept gains in welfare and revenue while preserving response quality.
- Motivation: Generative AI synthesizes responses dynamically, making advertising opportunities dependent on how content unfolds.Unlike retrieval and ranking systems, AI-native interfaces shape content around user intent during interaction.
- Motivation: Generated content can create advertising opportunities, so predefined-slot abstractions no longer fully capture when and how ads fit.An early generation choice may make a later brand mention natural or jarring.
- Method: LAMA embeds advertiser participation into token-level generation, jointly determining response content, advertising allocation, and payments.At each prefix, advertiser messages influence the next-token distribution and the evolving allocation process.
- Theory: LAMA satisfies Markov DSIC and IR and achieves near-optimal KL-regularized welfare with an optimality gap vanishing as regularization diminishes.The introduction also gives an additive welfare gap of at most βlog |N|.
- Implementation: A learning-based implementation reconstructs LAMA’s reports online from learned local signals and advertiser-specific root estimates.The serving procedure avoids explicitly reasoning over the full token tree and supports efficient operation over small advertiser candidate sets.
- Evaluation: On real-world search query splits, LAMA improves platform welfare and revenue while maintaining advertiser value and user-facing response quality.The evaluation compares LAMA with mechanisms allocating advertising before or after generation and with response-level aggregation.
2 Model
The paper models token-level advertising as a sequential mechanism in which reports, token generation, payments, and final allocation depend on evolving generation histories. It formalizes advertiser types and truthful behavior, and evaluates welfare using a KL penalty that balances advertiser value against deviation from organic generation.
- Sequential model: A user query initializes a sequential token-generation process, and each generated token deterministically extends the current state.Generation terminates when the state contains EOS or reaches maximum length L.
- Advertisers and allocation: Each advertiser has a private terminal reward function mapping terminal states to nonnegative values, and the paper focuses on a single-winner feasible allocation.The single-winner setting uses Λ = Δ(N), with the realized gross value given by λ_i r_i(ℓ(τ)).
- Sequential mechanisms: Advertisers submit local messages from feasible, history-dependent message spaces, while the platform uses reports and private history to choose tokens, payments, and allocation.Advertisers observe only the current public state; the platform observes the entire platform history.
- Reporting: The model distinguishes lightweight dynamic reports from direct mechanisms that require advertisers to submit entire reward functions.Lightweight reports are emphasized because full reporting may be slow and may reveal information about unrealized trajectories.
- Incentives: Markov DSIC requires truthful reporting to remain optimal after every reachable history following truthful prefix reports, while IR requires nonnegative expected utility from truthful initial participation.This sequential definition prevents advertisers from waiting for favorable prefixes before manipulating future reports.
- Welfare: KL-regularized social welfare combines gross advertiser value with a divergence penalty against the organic policy, with β controlling monetization versus content naturalness.The mechanism-design objective maximizes induced welfare for query-type pairs under the reference language-model policy.
3 Latent Advertiser Mixture Auction
LAMA treats generation as Bayesian sequential decision-making: advertiser reports induce token policies, while posterior updates connect token generation to final allocation. It provides incentive and efficiency guarantees for this sequential mechanism.
- Mechanism: LAMA samples a latent advertiser whose induced policy generates each next token, then updates allocation beliefs after observing the token.The final posterior determines the allocation and payments after generation terminates.
- Payments: LAMA uses entry fees and path payments based on changes in posterior-weighted continuation values.Advertisers are charged when realized tokens improve continuation prospects and subsidized when they worsen.
- Mechanism: Advertisers submit child-value reports constrained by a soft Bellman recursion checked against values stored in the platform ledger.This history-dependent feasibility condition lets the platform verify consistency online.
- Guarantees: LAMA is Markov DSIC and individually rational under Bellman-consistent message spaces and truthful soft-value reporting.The proof connects online deviations to globally Bellman-consistent direct reports, then applies direct DSIC and IR.
- Guarantees: The mechanism’s welfare loss is at most βlog |N| relative to the optimal KL-regularized benchmark and approaches first-best welfare as β→0+.Truthful LAMA maximizes the entropy-relaxed welfare objective.
4 Practical Implementation
The practical implementation reconstructs LAMA’s required reports with learned local advantages and root values. A shared advertiser-conditioned model serves these reports online with lightweight overhead while retaining the mechanism’s guarantees when reports are exact.
- Motivation: Individual advertisers may lack the feedback data and computation needed to report full continuation-value vectors, so the platform serves reports on their behalf.This platform-side service is analogous to autobidding.
- Value Decomposition: The platform decomposes reporting into local soft advantages, which capture value changes between prefixes, and a root value anchoring the absolute scale.The two outputs are assembled along the realized path to reconstruct required reports.
- Recovery: Under full comparison support, the learned local-advantage objective uniquely recovers the optimal token policy and truthful soft-value reports.The recovered child-value vector is exact at every reachable prefix.
- Serving Model: A single shared pretrained backbone produces advertiser-specific next-token logits and a scalar root value through advertiser identity and campaign conditioning.The deployed model shares parameters across advertisers while conditioning on creative, landing-page, and targeting information.
- Serving Model: At serving time, root values initialize the ledger and allocation posterior, while conditioned logits generate local advantages at each prefix.Because candidate advertiser sets are typically small, serving adds only a small constant-factor overhead over reference-answer generation.
- Implementation Guarantee: With exact served reports, winner-settled LAMA is outcome individually rational and weakly budget balanced in expectation.Non-winners pay zero and the winner receives non-negative realized utility.
5 Experiments
Experiments evaluate LAMA on three real-world commercial-search verticals against before-generation, after-generation, and response-level baselines. LAMA obtains the strongest overall results while preserving user-facing quality.
- Baselines: The evaluation compares LAMA with heuristic mechanisms that allocate before or after generation and with MOSAIC, a response-level aggregation mechanism.MOSAIC generates complete candidate responses, scores aggregate advertiser value, and selects a weighted-lottery outcome.
- Evaluation: The experiments measure platform welfare and revenue, advertiser value, and user-side response quality.Welfare subtracts the KL regularization term representing user-experience cost, while revenue is the winning advertiser’s payment.
- Main Results: LAMA achieves the highest mean platform welfare (0.5205), revenue (0.8305), advertiser value (0.8568), and user quality (66.5239).The evaluation covers Workout, Vacation, and Car query splits with three heterogeneous advertisers per vertical.
- Main Results: Compared with the allocate-after baseline, LAMA improves revenue while also gaining welfare and advertiser value.It matches the best-performing baseline on user-facing response quality.
- Conclusion: The results provide an initial proof of concept that token-level advertiser participation can improve monetization without sacrificing response quality.The paper presents this as initial evidence for generation-native advertising.
6 Related Work
Prior work has moved advertising from fixed slots toward context-dependent segments, positions, and generated responses. LAMA extends this progression by allowing advertiser influence at token-level granularity throughout generation.
- Prior Mechanisms: Classical advertising auctions allocate exogenously defined inventory such as sponsored-search or display slots.Later work incorporates cascade effects, externalities, richer click-through structures, and joint advertising-content optimization.
- Generative Advertising: Recent generative-interface mechanisms define opportunities over discourse segments, context-dependent position–creative pairs, or complete candidate responses.These approaches make advertising more adaptive to generated context but still operate on a preidentified segment, position, or response.
- LAMA’s Position: LAMA allows advertisers to influence generation at token-level granularity rather than allocating opportunities identified before or after generation.Its influence evolves with the current prefix, posterior beliefs, and continuation values.
7 Conclusion
The paper introduces generation-native advertising, with LAMA shaping advertising opportunities during token-level generation. It reports Markov DSIC and IR, bounded welfare loss, a practical implementation, and improved monetization while maintaining response quality.
- LAMA shapes advertising opportunities during generation rather than allocating only after they have formed.
- LAMA instantiates generation-native advertising at token level and provides a practical learning-based implementation.
- LAMA achieves Markov DSIC and IR with bounded welfare loss.
- Proof-of-concept experiments show improved monetization while maintaining user-facing response quality.
A Report-Model Training
This section identifies Algorithm 2 as the practical implementation of the two decomposed learning tasks described in Section 4.1.
- Algorithm 2 implements the two decomposed learning tasks from Section 4.1.
B Potential Business Format
LAMA can be applied through three representative single-winner business formats: an embedded hyperlink, a clickable response block, or a sponsored card alongside the response.
- LAMA’s single-winner setting supports a hyperlink embedded in the generated response, a clickable response block, and a sponsored card displayed alongside it.
C Proofs
The proofs establish LAMA’s soft dynamic-programming structure, allocation representation, incentive properties, and welfare guarantee. They also connect the mechanism to entropy-regularized allocation and expected budget balance.
- Soft Bellman structure: The optimal continuation policy and value satisfy a soft Bellman recursion in the finite-horizon token MDP.
- Allocation representation: Under Bellman-consistent reports, the online decision rule and terminal allocation admit a desirability-based posterior representation.
- Allocation representation: The joint allocation rule is the gradient of a convex soft potential and is therefore cyclically monotone.
- Incentives: Truthful reporting is DSIC and IR, with Markov DSIC extending incentive compatibility to reachable future states.
- Welfare: The terminal allocation uniquely optimizes an entropy-regularized allocation problem because Shannon entropy is strictly concave.
- Payments: LAMA is weakly budget balanced in expectation over token-generation randomness and the terminal winner draw.
D Additional Experiment Results
The appendix documents experimental settings, generated-response case studies, latency, impression-value calibration, and bid-offset monotonicity diagnostics for LAMA.
- Experimental setup: The experiments cover three verticals—Workout, Vacation, and Car—with heterogeneous advertisers and advertiser metadata summarized in Table 2.
- Generated responses: Case studies show the winning advertiser in bold and identify the response span allocated to its hyperlink.This presentation is used for the Workout, Vacation, and Car verticals.
- Efficiency: Latency ratios between LAMA and reference-answer generation are reported across verticals and are close to the theoretical prediction (|N| + 1).
- Report-model diagnostics: Table 9 and Figure 4 report impression-value prediction and calibration performance across verticals, with values defined as click-through rates multiplied by unit cost per click.The current Qwen3-14B plus LoRA fine-tuning setup yields reasonably good recovery accuracy.
- Bid-perturbation diagnostics: Figure 5 evaluates monotonicity by adding offsets from −0.3 to +0.3 to one target advertiser’s truthful value while holding other bids and advantage settings fixed.The targets are CARFAX for Car, ClassPass for Workout, and Tripadvisor for Vacation.