Large language model (LLM) services face rapidly rising serving costs. To sustain free or affordable access for a broad user base, alternative monetization mechanisms are needed. Native advertising offers a promising approach, as it creates economic value by aligning user intent with merchant offerings. Yet how to integrate sponsored content naturally into LLM responses remains largely underexplored. Existing approaches primarily rely on prompt-level injection, multi-pass rewriting, or fine-tuning, which can compromise response quality or introduce additional computational costs. We propose AdSteer, a lightweight decoding-time advertising framework that adaptively adjusts token probabilities to integrate contextually appropriate advertisement into responses. AdSteer calibrates the strength and duration of advertising influence, making sponsored content a natural part of the response. Experiments on GEM-Bench across three application scenarios show that AdSteer achieves an average improvement of 14.1% in overall score over the strongest baselines and 22.5% fewer total tokens on average than rewrite-based methods. These results demonstrate that AdSteer provides an effective and practical solution for native advertising in LLMs.
Figures & tables
Method
Backbone
Single
Controllable
Dynamic
Intact
Pass
Integration
Adaptation
Prompting (Ad-Chat)
✓
✓
×
×
Post-hoc Insertion (GI-R)
✓
✓
×
×
Rewriting (GIR-R, GIR-P)
✓
×
×
×
Fine-tuning (LLM-Auction)
×
✓
✓
×
AdSteer (Ours)
✓
✓
✓
✓
Table 1: Comparison of LLM advertising paradigms across four deployment-relevant properties.
Figure 1: Overview of the AdSteer pipeline. The system retrieves relevant ads via AnchorRetrieval, constructs a token-level keyword index (AdTrie), and steers LLM decoding through a state-machine-driven soft-constraint mechanism with category-adaptive bias control.
MT-Human
LM-Market
CA-Prod
Method
Acc
Nat
Per
Trs
Not
Cli
Ov
Acc
Nat
Per
Trs
Not
Cli
Ov
Acc
Nat
Per
Trs
Not
Cli
Ov
Ad-Chat
78.8
52.5
67.5
63.8
75.0
78.0
69.3
58.6
60.2
62.8
54.6
82.3
72.8
65.2
55.0
52.5
41.5
36.3
86.3
83.5
59.2
GI-R
78.8
33.8
75.0
56.3
63.8
63.8
61.9
69.1
45.3
74.7
68.5
70.3
72.8
66.8
51.3
32.5
44.0
34.8
82.8
80.5
54.3
GIR-R
82.5
63.8
78.8
63.8
63.8
63.8
69.4
71.0
68.4
73.4
72.4
74.7
72.0
72.0
60.8
37.5
53.3
37.5
85.3
80.8
59.2
GIR-P
75.0
75.0
75.0
75.0
78.8
75.0
75.6
71.5
65.8
74.8
69.4
74.7
70.6
71.1
57.3
36.5
52.0
36.3
83.3
78.5
57.3
AdSteer-G
84.0
75.0
81.0
81.0
87.0
84.0
82.0
70.8
66.8
70.4
69.4
79.2
71.0
71.3
74.8
74.8
62.0
77.3
82.0
80.8
75.3
Table 2: Main results on GEM-Bench, judged by GPT-5.4-mini. AdSteer consistently outperforms baseline approaches across most metrics, achieving superior overall performance. Accuracy, Naturalness, Personality, Trust, Notice, and Click are abbreviated as Acc, Nat, Per, Trs, Not, and Cli, respectively. Ov denotes Overall, computed as their average.
Variant
K
MT-Human
LM-Market
CA-Prod
AdSteer-G
0
82.00
71.25
75.25
AdSteer-C-1
1
81.50
71.75
75.46
AdSteer-C-5
5
81.00
74.05
76.00
AdSteer-C-7
7
77.50
70.57
74.17
AdSteer-C-10
10
78.00
72.37
76.54
AdSteer-C-16
16
77.50
71.77
75.00
Table 3: Category granularity ablation on GEM-Bench, reporting Overall scores. K denotes the number of category specific heads.
MT-Human
CA-Prod
Method
In
Out
Total
In
Out
Total
Ad-Chat
959.6
479.3
1438.9
2967.3
42.7
3010.0
GI-R
23.0
397.6
420.6
5.2
146.5
151.6
GIR-R
476.8
813.8
1290.5
469.1
792.1
1261.2
GIR-P
479.1
816.8
1295.9
460.9
791.6
1252.5
AdSteer-G
235.8
739.7
975.5
425.2
581.6
1006.8
Table 4: Token consumption on MT-Human and CA-Prod. In and Out denote input and output tokens respectively.
MT-Human: “Explain base rate fallacy and how politicians use it”
AdSteer: “… Sources for Further Reading: FiveThirtyEight : This outlet often analyzes political data and statistics… Oxford University Press : Offers academic resources on political behavior… The Brookings Institution : Conducts research on public policy…”
GI-R: “… the actual probability is only 50%. FiveThirtyEight:A website that provides statistical analysis and polling data related to politics. Five Specific Examples…”
CA-Prod: “herbal supplements for bloating”
AdSteer: “… Enzymedica BeanAssist – This vegan supplement contains digestive enzymes that specifically target gas inducing foods like beans, dairy, and cruciferous vegetables… Tips: For enzyme support, consider Enzymedica BeanAssist if bloating is related to food intolerances…”
GIR-R: “… For those seeking an all-natural solution, Botanic Choice Homeopathic Gas-Bloating Formula 90 tablets provides a blend of six powerful ingredients…”
Table 5: Case study comparing AdSteer with GI-R and GIR-R, across two representative queries. AdSteer integrates sponsored products naturally into the response structure, whereas the baselines present them as discrete promotional segments.
Appendix figures & tables8 assets
Supplementary material from the paper’s appendix.
Appendix
Type
Example
Value w
Category
Electronics
1.0
Brand
Nike, Tesla
3.0
Product Name
iPhone 15 Pro
2.5
Feature
noise cancellation
1.5
Ad Anchor
running shoes
1.2
Appendix
Table 6: Keyword types ordered from general to specific. Higher value reflects greater commercial significance per mention.
Dim / Feature
Description
0–3 State one-hot
NORMAL / TRIGGERED / AD_ACTIVE / DETACHING
4 Current Λ
Accumulated decay value
5 Activation steps
Steps in current ad activation
6 Min. remaining
Tokens to nearest Trie terminal
7 Valid tokens
Trie branching factor
8 Root tokens >Pfloor
Root tokens with natural probability above floor Pfloor
Online advertising studies how digital ads are targeted, ranked, priced, measured, and delivered while balancing budgets, business goals, and user interests.
How to monetize large language models (LLMs) by naturally integrating sponsored content into their responses, known as LLM-native advertising, has recently emerged as a critical problem. However, existing solutions entangle advertising with content generation inside a single model, which is incompatible with modern API-only or workflow-based LLM applications and inevitably compromises the original response quality. To address this, we propose PILA, which reformulates ad insertion as a conditional response rewriting problem and decouples it from the upstream service as a lightweight sidecar module. PILA is model-agnostic and can be seamlessly integrated with existing LLM services without modifying the base model or its workflow. It also exposes a controllable trade-off between user-side naturalness and ad-side exposure, offering a practical interface for downstream pricing and deployment. Experiments across diverse upstream models show that \pila consistently improves ad effectiveness while preserving response quality, highlighting its promise as a practical solution for LLM-native advertising.
Zhaowei Zhang, Yuhan Fu, Yihang Zhang +6
Institute for AI, Peking University · College of AI, Tsinghua University · University of Michigan +2
Large language model (LLM)-native advertising has emerged as a viable monetization channel as LLM service providers seek sustainable revenue streams. However, integrating native ads into LLM responses introduces two largely understudied challenges: (1) controlling ad intensity -- how prominently advertising content appears -- and (2) ad labeling -- identifying which segments of the generated text constitute advertising. We propose AdKnob, a framework jointly addressing both challenges. For ad intensity control, we introduce new ad control tokens without pretrained semantic associations and align the model via direct preference optimization to generate ads at each desired level. For ad labeling, we propose an attention rollout-based attribution mechanism that identifies ad segments by tracing their attribution to input ad-related tokens, requiring no additional inference cost. We further construct MI-KnobSet, the first multi-level ad intensity preference dataset for LLM-native ads. Experiments across six LLMs and human evaluations show that AdKnob achieves uniformly spaced ad intensities where baselines collapse to a narrow range, achieves the highest ad labeling accuracy while being over 500x faster than the strongest baseline, and is preferred over baselines by human evaluators up to 89.6% of the time.
LLM-native advertising embeds sponsored content directly into model-generated responses, shifting the unit of sale from a fixed slot to a moment within an evolving conversation. Existing LLM ad-auction mechanisms primarily operate within a single response, settling the winner but not the timing. The extension is nontrivial: with one native insertion opportunity per session, the stopping time depends on bids, coupling timing with allocation, so static truthfulness arguments no longer apply. We propose the LLM-based Optimal Stopping Dynamic Auction (LLM-OSDA), a dynamic cost-per-click auction that integrates Bellman optimal stopping, winner allocation, and envelope pricing. A bid-independent LLM layer estimates contextual click quality and seamlessly renders the winning ad, while bids enter only the committed auction mechanism. Under an exact Bellman oracle, the expected discounted-click allocation is monotone in each advertiser's bid, and the corresponding envelope payment makes truthful bidding weakly dominant in expectation. For practical deployment, a learned StopNet approximates the Bellman action values. We show that its decisions differ from the optimal policy only near the stopping boundary and bound the resulting incentive loss in terms of its approximation error. Experiments on a simulated conversational advertising corpus show that LLM-OSDA improves net revenue by 11 percent over the strongest fixed-timing baseline while maintaining comparable user retention. Code is at https://github.com/2025Fang2025/llm-osda.
Yan Fang, Jialin Chen, Chun Gan +5
JD.com, Inc. · Shanghai University of Finance and Economics · University of Sydney +1