AdKnob gives AI-generated advertising an intensity dial—and a way to mark it

Advertising inside an AI-generated answer creates a different disclosure problem from a banner beside it. Promotional material can appear as a recommendation, a product description or a seemingly ordinary explanation. In the preprint AdKnob: Ad Intensity Control and Labeling for LLM-Native Advertising, Woo Jae Kim and colleagues at KAIST propose training language models both to regulate that promotional emphasis and to identify the sentences carrying it.

The distinction matters: generating an advertisement, controlling its prominence and labeling it are separate tasks. The authors argue that existing approaches largely concentrate on inserting advertising, leaving its intensity and disclosure less explicitly controlled. AdKnob combines those latter objectives in one training framework.

Teaching the model what the dial means

AdKnob introduces special vocabulary tokens corresponding to advertising levels. In the main experiments, four levels run from the weakest to the strongest promotion. The selected token is appended to the user’s request, while the brand and its promotional message appear in the system prompt.

Rather than relying on familiar words such as “low” or “high,” the researchers give these new tokens embeddings initialized around the pretrained vocabulary’s mean, with added randomness. Their purpose is to avoid importing existing linguistic associations. In the paper’s comparisons, models trained with ordinary descriptors sometimes insert those words awkwardly into answers rather than using them solely as control signals.

The researchers then use direct preference optimization, a method that trains a model to favor selected responses over rejected alternatives. Crucially, the selection rewards both the intended advertising intensity and a useful answer. Rejected examples separately represent poor response quality, incorrect intensity, or both. The aim is not simply to make stronger ads, but to preserve relevant information while changing promotional emphasis.

For this training, the team constructs MI-KnobSet from an existing native-advertising dataset. GPT-4o helps turn its material into conversational questions, promotional descriptions and candidate answers. The resulting dataset contains 3,946 training instances and 913 test instances across ten topics. Each training instance produces 12 preference pairs across the four levels.

Finding promotional sentences

Disclosure uses a different signal: how strongly generated text depends on the brand and advertising copy supplied to the model.

During training, the researchers compare token probabilities with and without those advertising inputs. The differences provide automatically generated supervision for an additional attribution loss. That loss trains attention rollout—a technique for tracing influence through the model’s attention layers—to associate advertising inputs with promotional output.

At generation time, AdKnob uses attention weights already computed by the model, averages attribution scores across each sentence and labels sentences exceeding a threshold. This avoids extra model passes for labeling, although the attribution calculation itself has a measured runtime. On Qwen3-4B, the paper reports 3.3 milliseconds per response, versus 1,785 milliseconds for its BERTScore labeling baseline. The more-than-500-fold speed comparison concerns labeling, not the time needed to generate an entire answer.

What the evaluations show

Across six language models, the authors report a wider, more evenly spaced range of advertising intensities than the tested alternatives. Human ratings also distinguish the levels: average perceived intensity rises from 1.30 to 3.66 on a four-point scale. Response quality ratings decline from 3.58 to 3.10 as promotion increases, making the quality–intensity tradeoff visible even within the favorable results.

In forced-choice comparisons using AdKnob at Level 3, crowd workers preferred its responses over the four baselines at rates ranging from 75.0% to 89.6%. Reported click preference reached 79.8%. Those are judgments about which advertisement participants would be more likely to click, not observed click-through rates.

Human-annotated sentence labels provide another important check. AdKnob achieves recall of 0.86, precision of 0.58 and F1 of 0.69, outperforming the compared labeling methods on F1. Those figures also show that disclosure remains imperfect: the system misses some promotional sentences and labels some sentences that annotators did not classify as advertising.

A technical control, not a deployment verdict

The authors explicitly limit the setting. AdKnob assumes an advertised brand relevant to the request; choosing that brand and deciding when to advertise remain separate allocation problems. It also requires fine-tuning and access to internal attention weights, so it is not directly usable through closed-model APIs lacking those capabilities.

Most importantly, the evaluation is offline, without live click-through or retention measurements. The preprint offers evidence for a combined control-and-labeling mechanism—not a demonstrated commercial outcome. Its central contribution is making promotional strength and disclosure explicit parts of model training rather than leaving both to emerge from advertising instructions.