AdKnob: Ad Intensity Control and Labeling for LLM-Native Advertising

AI summary

Overview

Research area: LLM alignment and AI-generated content — specifically, monetization of large language model assistants through native advertising (ads woven directly into model-generated text rather than shown as banners).

Technical level: Intermediate. The paper uses standard alignment machinery (direct preference optimization, LoRA fine-tuning, attention rollout), but the motivation and results are accessible without deep background.

Scope: The paper proposes AdKnob, a framework that gives one LLM a continuous dial over how strongly ads appear in its responses and, at the same time, automatically flags which sentences of a response are advertising.

What This Paper Is About

LLM assistants are increasingly paid for by embedding brand promotions directly into free-form answers, which makes ads harder for users to spot and harder for providers to control. Existing methods can inject ads but cannot dial how prominent they are, and they leave no reliable way to say which parts of the generated text are ads — something regulators such as the FTC, IAB, and the EU Digital Services Act require. AdKnob addresses both gaps in a single aligned model: a requested ad intensity level as input, and sentence-level ad labels as a by-product of generation.

Key Contributions

1. Ad control tokens for intensity control. The authors add N special tokens (one per intensity level, [AD_LV_n]) to the model vocabulary, initialized with random values added to the mean of all pretrained token embeddings, so they carry no prior linguistic meaning. DPO then learns a clean mapping from each token to a target ad intensity.

2. Attention-rollout ad labeling with no extra inference cost. Ad labeling is reframed as an attribution problem: a sentence is advertising if its tokens depend strongly on the input brand and ad-copy tokens. Attention rollout traces this dependence using only attention weights from the original forward pass.

3. An ad-attribution loss. Because DPO alone does not establish strong attribution, an auxiliary loss supervises attention flow from ad-related input tokens to generated ad segments, using ablation-based pseudo-labels (the drop in a token's log-probability when the brand and ad copy are removed from the prompt) as the target.

4. MI-KnobSet, the first multi-level ad intensity preference dataset. Built by extending the Webis Generated Native Ads 2024 dataset into a conversational setting with GPT-4o-curated ad copies and user questions, yielding 3,946 training and 913 test instances across four intensity levels.

Main Findings

• Intensity levels are actually distinguishable. Across four base models shown in the intensity figure, AdKnob (L1–L4) produces a wide, uniformly spaced range of advertisement scores, while baselines collapse into a single or narrow intensity band. Natural-language prompt levels (NL-prompt) overlap heavily, confirming that prompt wording alone cannot control intensity.

• Human perception matches the dial. In a user study with 913 participants on 913 test responses, mean human-rated ad intensity rose monotonically from 1.30 at Level 1 to 3.66 at Level 4 on a 1–4 Likert scale, with all adjacent-level differences statistically significant (Mann–Whitney U, p < 0.001).

• The advertisement score tracks human judgment. The paper's s_ad metric correlates with human-perceived ad intensity at Spearman ρ = 0.91.

• Quality degrades only modestly as intensity rises. Human-rated response quality goes from 3.58 at Level 1 to 3.10 at Level 4, which the authors attribute to preference pairs that explicitly reject low-quality responses.

• Best ad labeling accuracy across all six models. Against ablation-based pseudo-labels (δ* = 1.0, τ = 0.005), AdKnob reaches the best AUC and F1 on every model — for example AUC 0.77 / F1 0.76 on Qwen3-4B, 0.76 / 0.75 on Phi-4-Mini (3.8B), 0.79 / 0.76 on Qwen2.5-14B, and 0.77 / 0.62 on Phi-4 (14B).

• Real-time labeling. Labeling takes about 3 ms per response on Qwen3-4B (3.3 ms), versus roughly 1,800 ms for BERTScore and 179 ms for AttnLRP — over 500× and 50× faster respectively. The LLM-based Marker baseline is consistently low-recall, meaning models struggle to self-identify their own ad content.

• Human-annotated labeling confirms the gain. Against human sentence labels, AdKnob scores AUC 0.86, precision 0.58, recall 0.86, and F1 0.69, versus BERTScore at 0.82 / 0.52 / 0.71 / 0.60 and AttnLRP at 0.47 / 0.21 / 0.63 / 0.31. High recall matters because missed disclosures carry legal and trust risk.

• Pseudo-labels are trustworthy. The ablation-based pseudo-labels agree with human annotations at Cohen's κ = 0.84 and F1 = 0.88.

• Humans prefer AdKnob's responses. A/B tests at Level 3 against four baselines give user-preference win rates of 89.6% (vs. Two-step), 81.3% (vs. Prompt-based), 79.2% (vs. JU-NLP), and 75.0% (vs. NL-prompt); click-likelihood win rates reach up to 79.8%. The most cited reasons are helpfulness (80%) and ad disclosure (67%) for user preference, and non-intrusiveness (76%) and relevance (67%) for click preference.

• Ablations justify the design. Replacing control tokens with qualitative descriptors ("low", "medium-low"...) causes descriptor words to leak into responses and hurts quality; numeric descriptors ("level 1"–"level 4") fail to give uniformly spaced intensity. Using all three rejected-response criteria (quality failure, control failure, both) beats any single criterion. γ = 10 gives the best balance between intensity range and response quality.

Methodology in Plain English

The setup involves a system prompt carrying a brand and a short ad copy, plus a user question. The goal is a response that answers the question while promoting the brand to a requested degree.

Controlling intensity. Rather than telling the model "be subtle" or "be aggressive" in words, the authors create brand-new vocabulary tokens with no learned meaning, one per intensity level, and append the relevant one to the user prompt. Because these tokens are semantically blank, the model cannot fall back on prior associations and must learn what each one means purely from training. Training uses direct preference optimization: for each prompt and level, the model is shown a preferred response and several dispreferred ones and learns to favor the former. Preferred responses are chosen by a score that rewards hitting the target advertisement score while keeping response quality high. Dispreferred responses are picked to isolate three distinct failure modes — bad quality, wrong intensity, and both at once — so the model receives an unambiguous learning signal rather than a muddled one.

Measuring ad intensity and quality. The advertisement score combines three things: what fraction of sentences mention the brand, the sentiment of those mentions (via DistilBERT), and how semantically close the response is to the ad copy (cosine similarity of LLaMA-3-8B hidden states). Brand sentiment is multiplied by the hit rate because a negative brand mention should not count as high intensity. The quality score comes from an LLM-as-a-judge setup averaged across LLaMA-3-8B-Instruct, Qwen3-4B-Instruct, and GPT-4o-mini, scoring request relevance and information preservation relative to a no-ad reference response.

Labeling the ads. Since ads are paraphrased rather than copied verbatim, string matching fails. The authors instead ask: which output tokens would have been unlikely if the brand and ad copy had not been in the prompt? Computing this directly requires many extra forward passes, so it is used only as training-time supervision. During training, an extra loss pushes the model's internal attention to flow from the brand and ad-copy input tokens toward genuinely ad-related output tokens. At inference, the model simply averages these attention-flow scores per sentence and flags any sentence above a fixed threshold.

Evaluation. Six models (Qwen3-4B-Instruct, Phi-4-Mini-Instruct, LLaMA-3.2-3B-Instruct, Qwen2.5-14B-Instruct, Phi-4, LLaMA-3.1-8B-Instruct) are fine-tuned with LoRA (β = 0.01, λ = 1.0, γ = 10) on a single NVIDIA H200 GPU, generating up to 500 tokens, and compared against four generation baselines and four labeling baselines, plus MTurk human studies.

Why This Matters

Impact on research. The paper reframes ad labeling from a post-hoc classification task into an attribution problem that rides along on attention weights already computed during generation — a template that could apply to other disclosure problems in generated text (sponsored claims, safety-relevant statements, provenance). It also offers the first multi-level ad intensity preference dataset, giving future work a controllable benchmark rather than a binary ad/no-ad setup.

Real-world applications:

• Subscription tiers. A provider could offer ad-free premium tiers and progressively lighter ad tiers on cheaper plans, all served by one model.
• Contextual governance. Suppress ads entirely for urgent medical or safety queries while allowing higher intensity in shopping contexts.
• Regulatory compliance. Automatic sentence-level disclosure satisfies FTC, IAB, and EU Digital Services Act expectations without a separate auditing model in the serving path.
• Advertiser reporting and billing. Knowing exactly which sentences carried the ad enables impression measurement and intensity-based pricing.

Industry relevance. The ~3 ms labeling cost is the headline for practitioners: compliance labeling that adds seconds per response is not deployable at chat latency, whereas one that reuses existing attention weights is essentially free. The framework is model-agnostic across both 3–4B and 14B models, which matters for cost-sensitive serving.

Future Directions

• Calibration and governance at scale. The paper fixes thresholds (τ = 0.005, δ* = 1.0) and finds τ generalizes across models, but how such a dial should be governed, audited, and defended to regulators is not explored.
• Beyond sentence granularity. The authors deliberately average over sentences because paraphrased ad content is distributed rather than localized; finer-grained or partial-sentence disclosure, and cases where an ad message spans multiple sentences, are open.
• Human perception of intermediate levels. The user study validates four levels; whether users can reliably distinguish more granular levels, and where the trust cost of higher intensity actually bites, is not reported.
• Generalization beyond the curated data. MI-KnobSet is derived from one native-ad dataset extended with GPT-4o; behavior on other domains, languages, and advertiser styles is untested in this paper.

Target Audience

Researchers and engineers working on LLM alignment, preference optimization, and controllable generation; applied scientists at LLM assistant or search providers building ad-supported products; and trust-and-safety, policy, or compliance teams who need to reason about disclosure of sponsored content in generated text. Readers focused on advertising auction mechanics or click-through modeling will find less here — the paper is about generation control and disclosure, not ad targeting or bidding.

Authors’ abstract

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.