LLMAdBench: A Human Preference Benchmark for Advertising in LLM Responses

LLMAdBench is a controlled, human-annotated benchmark for studying where advertisements should be placed in LLM-generated responses. Given a user conversation, a base response, and a matched advertisement, the benchmark compares response pairs that differ in advertisement placement while holding the underlying conversation, answer, advertised entity, core advertising message, and disclosure condition fixed.

The benchmark contains 258 source conversations, each expanded into four placement candidates—front, early, late, and end—and all six unordered placement pairs. The same materials are evaluated under two disclosure conditions:

• Labeled: the advertisement is explicitly marked as [Sponsored: <ad text>].
• Unlabeled: the advertisement is integrated without an explicit sponsorship marker.

Human annotators compare Version A and Version B and select A, B, or Tie along six dimensions. In total, the source data contain 1,548 pairwise comparisons per disclosure condition and 18,576 dimension-level human judgments.

Links

• Dataset: ruiai-work/LLMADBENCH
• Paper: arXiv:2609.32533

Supported Tasks

LLMAdBench supports research on:

• three-way human-preference prediction (A, B, or T);
• pairwise ranking of advertisement placements;
• preference-model and LLM-judge training;
• robustness to A/B presentation-order swaps;
• the effects of advertisement placement and sponsorship disclosure;
• trade-offs among user experience, advertisement visibility, and stated click likelihood.

For benchmark evaluation, report exact three-way accuracy separately for the labeled and unlabeled conditions. Because rows derived from the same source comparison are related, use the provided train/test files rather than creating a row-level random split.

Benchmark Results

We report the held-out exact-match accuracy (%) for three-way A/B/T prediction. The first eight models are evaluated as zero-shot judges. Pre-DPO and Post-DPO refer to Qwen3-8B before and after five epochs of Direct Preference Optimization on the corresponding disclosure condition.

Condition	GPT-5.4	Claude Opus 4.6	Gemini 3.1 Pro	Grok 4.1	Kimi K3	DeepSeek V4 Pro	Llama 4 Maverick	Mistral Large 3	Pre-DPO	Post-DPO
Labeled	47.85	46.96	48.14	46.31	47.23	44.00	44.51	46.39	40.25	50.54
Unlabeled	50.32	48.74	49.92	47.52	49.87	46.18	48.09	48.76	36.52	55.76

Evaluation Dimensions

Dimension	Annotation question
Naturalness of Flow	Which version reads more smoothly and feels less disruptive?
Ad Visibility	In which version is the advertisement easier to notice? Visibility is a neutral perceptual judgment.
Content Helpfulness	Which version better preserves the usefulness of the non-advertising content for answering the user?
Click Likelihood	If required to click one advertisement, which version's advertisement would the annotator be more likely to click?
Appropriateness	Which version feels more trustworthy and less manipulative, including whether the advertisement is suitable, transparent, and not overly pushy?
Overall Preference	Which version does the annotator prefer, all things considered?

Click likelihood represents stated behavioral intention, not observed clicking behavior.

Dataset Files

Source comparison files

File	Rows	Description
Labeled.parquet	1,548	Human comparisons with explicit sponsorship disclosure.
Unlabeled.parquet	1,548	Human comparisons without explicit sponsorship disclosure.

Each row contains one placement pair and all six human judgments.

Preference-model files

Configuration	Train	Test	Total
Labeled	23,530	3,714	27,244
Unlabeled	21,381	3,716	25,097

These files transform the source comparisons into dimension-level A/B/T preference examples, including A/B-swapped variants and preference-versus-tie examples. The split is constructed at the original comparison level before augmentation, using an 80/20 split, so related augmented rows do not cross train and test.

Loading the Dataset

from datasets import load_dataset

labeled = load_dataset("ruiai-work/LLMADBENCH", "labeled")
unlabeled = load_dataset("ruiai-work/LLMADBENCH", "unlabeled")

print(labeled["train"][0])

To load the source comparison tables directly:

from datasets import load_dataset

source_labeled = load_dataset(
        "parquet",
        data_files="hf://datasets/ruiai-work/LLMADBENCH/Labeled.parquet",
        split="train",
)

Data Schema

Source comparison schema

The files Labeled.parquet and Unlabeled.parquet share the following fields:

Field	Type	Description
original_ad	string	Original matched advertisement sentence or brief.
context	string	Previous conversation context, when available.
question	string	Current user request.
original_response	string	Base assistant response before advertisement insertion.
rewritten_ad_a	string	Advertisement wording used in Version A.
rewritten_ad_b	string	Advertisement wording used in Version B.
response_a	string	Candidate response shown as Version A.
response_b	string	Candidate response shown as Version B.
question_naturalness	string	Naturalness annotation prompt.
answer_naturalness	string	Human A/B/T judgment for naturalness.
question_visibility	string	Visibility annotation prompt.
answer_visibility	string	Human A/B/T judgment for visibility.
question_helpfulness	string	Helpfulness annotation prompt.
answer_helpfulness	string	Human A/B/T judgment for helpfulness.
question_click_likelihood	string	Click-likelihood annotation prompt.
answer_click_likelihood	string	Human A/B/T judgment for stated click likelihood.
question_appropriateness	string	Appropriateness annotation prompt.
answer_appropriateness	string	Human A/B/T judgment for appropriateness.
question_overall_preference	string	Overall-preference annotation prompt.
answer_overall_preference	string	Human A/B/T overall judgment.

Preference-model schema

The files labeled_{train,test}.parquet and unlabeled_{train,test}.parquet share the following fields:

Field	Type	Description
split	string	train or test.
question_id	string	Rubric question identifier (Q1–Q6).
dimension	string	Human-readable evaluation dimension.
dimension_instruction	string	Instruction defining the dimension.
ab_variant	string	orig for the original presentation or swap for the A/B-swapped presentation.
pair_type	string	Transformation type: pref_vs_alt, pref_vs_tie, or tie_vs_alt.
gold	string	Preferred completion label: A, B, or T.
rejected	string	Rejected completion label: A, B, or T.
previous_context	string	Previous conversation turns, or None.
user_question	string	Current user request.
version_a	string	First candidate response.
version_b	string	Second candidate response.
prompt	string	Complete judge prompt requesting a JSON choice.

The expected judge output is exactly one of:

{"choice": "A"}

{"choice": "B"}

{"choice": "T"}

Dataset Construction

Source data and matching

LLMAdBench combines conversational contexts from WildChat with advertisement material derived from the Webis Generated Native Ads 2024 dataset. The processing unit is one WildChat conversation turn. The current user question and assistant response are embedded with Qwen3-Embedding-0.6B and matched to candidate advertisements by cosine similarity. Up to five candidates above a similarity threshold of 0.65 are retained, after which Qwen3-8B filters irrelevant matches.

Human reviewers then curate the candidate pool, removing non-English, malformed, incomplete, meaningless, overly specialized, or otherwise unsuitable records. At most one advertisement candidate is retained for each conversation turn.

Placement candidates

Each retained conversation-advertisement pair produces four candidates:

1. Front: before the base response;
2. Early: at the earlier of the two selected internal sentence boundaries;
3. Late: at the later selected internal sentence boundary;
4. End: after the base response.

The two internal boundaries are selected by a fixed embedding-based continuity score. Qwen3-8B rewrites only the advertisement text to fit its local context while preserving the advertised product, brand, or entity and the core advertising information. The substantive base answer remains fixed.

Human annotation and quality control

Candidate order and task order were randomized. Annotators answered all six rubric questions with Version A, Version B, or Tie.

Quality control used two attention checks with verifiable answers, including an advertisement-placement check and a text-similarity check. Records failing either check were recollected in a second annotation round; replacement annotations passed both checks.

Intended Uses

The dataset is intended for research and evaluation involving advertisement placement in LLM responses, human-preference modeling, pairwise ranking, LLM judges, disclosure effects, and preference optimization.

It is not intended to:

• endorse any advertised product, brand, or claim;
• serve as evidence that an advertisement is truthful, safe, or legally compliant;
• predict real-world clicks, purchases, or conversion rates;
• justify undisclosed or deceptive advertising;
• replace user studies, legal review, or platform-specific safety evaluation.

License

LLMAdBench is released under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Users are also responsible for respecting the terms and licenses of the underlying WildChat and Webis-derived source materials.

Citation

If you use LLMAdBench, please cite:

@article{ai2026llmadbench,
    title   = {LLMAdBench: A Human Preference Benchmark for Advertising in LLM Responses},
    author  = {Ai, Rui and Liu, Yuqing and Qiu, Sitao and Qiao, Yun and Wang, Yuhan and Wang, Jessica Xiwen and Yang, Yiqi and Huang, Lihong and Sun, Ruiyao and Zhang, Kaifeng and Ding, Shengze and He, Jiaqi and Wang, Xinman and Gao, Tianhao and Qin, Jimmy and Lin, Jianghao and Wang, Chonghuan},
    journal = {arXiv preprint arXiv:2609.32533},
    year    = {2026},
    url     = {https://arxiv.org/abs/2609.32533}
}

Authors

Rui Ai, Yuqing Liu, Sitao Qiu, Yun Qiao, Yuhan Wang, Jessica Xiwen Wang, Yiqi Yang, Lihong Huang, Ruiyao Sun, Kaifeng Zhang, Shengze Ding, Jiaqi He, Xinman Wang, Tianhao Gao, Jimmy Qin, Jianghao Lin, and Chonghuan Wang.