Paper deep dive
A Systematic Investigation of Bias in Large Language Models for Advertising Relevance
Weiwei Wang, Yinchuan Xu, Jialu Gao, Youkow Homma, Jian Jiao
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/13/2026, 3:34:59 AM
Summary
This paper presents a systematic investigation into the fairness of Large Language Models (LLMs) when used for advertising relevance judgments. Using a counterfactual framework, the authors evaluate GPT-4o and a fine-tuned Qwen-7B model for biases related to advertiser identity (popularity), input language, and demographic wording. The study finds that both models exhibit sensitivity to these factors, with well-known advertisers receiving more favorable judgments and demographic terms triggering stereotype-consistent patterns. The paper also explores mitigation strategies, including company name masking during inference and training data rebalancing, finding that effectiveness depends on query intent and label distribution.
Entities (11)
Relation Signals (11)
Weiwei Wang â affiliatedwith â Microsoft
confidence 99% ¡ Affiliation: Microsoft
Qwen-7B â exhibitsbias â Advertiser Identity Bias
confidence 95% ¡ For both models, changing the advertiser identity or input language can alter the relevance assessment.
GPT-4o â exhibitsbias â Advertiser Identity Bias
confidence 95% ¡ For both models, changing the advertiser identity or input language can alter the relevance assessment.
Qwen-7B â exhibitsbias â Language Bias
confidence 94% ¡ Semantically equivalent inputs in different languages also lead to different outcomes in both models.
GPT-4o â exhibitsbias â Language Bias
confidence 94% ¡ Semantically equivalent inputs in different languages also lead to different outcomes in both models.
Qwen-7B â exhibitsbias â Demographic Bias
confidence 93% ¡ Selected demographic comparisons also show patterns consistent with common stereotypes
GPT-4o â exhibitsbias â Demographic Bias
confidence 93% ¡ Selected demographic comparisons also show patterns consistent with common stereotypes
Job Platform A â receivesfavorablejudgment â GPT-4o
confidence 92% ¡ Job Platform A has a lower Bad DR than Job Platform B... indicating that replacing the advertiser with Job Platform A produces more favorable GPT-4o judgments
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Large language models (LLMs) are increasingly used to judge how well an advertisement matches a query, but the fairness of these judgments has received limited attention. We conduct a systematic study of fairness in relevance judgments made by LLMs for queries and advertisements. Our counterfactual framework examines the effects of advertiser identity and possible popularity, input language, and demographic wording. We study GPT-4o as a categorical relevance judge and a Qwen-7B model trained specifically for relevance prediction. The advertiser and language experiments use query and advertisement pairs sampled from real advertising logs. Controlled synthetic queries are used to study demographic associations in employment, housing, and credit. For both models, changing the advertiser identity or input language can alter the relevance assessment. Selected demographic comparisons also show patterns consistent with common stereotypes, particularly those involving gender and occupation. We further study mitigation during model inference and training. The results indicate that its effectiveness depends on whether advertiser information is relevant to the query and how advertiser labels are distributed in the training data. These findings can help advertising practitioners identify fairness risks and develop suitable mitigation methods for LLM relevance systems.
Tags
Links
- Source: https://arxiv.org/abs/2610.07544v1
- Canonical: https://arxiv.org/abs/2610.07544v1
Trouble viewing inline? Open PDF directly â
Full Text
44,327 characters extracted from source content.
Expand or collapse full text
FandolSong-Regular.otf A Systematic Investigation of Bias in Large Language Models for Advertising Relevance Weiwei Wang â thanks: Correspondence to: weiwewang@microsoft.com. Affiliation: Microsoft Yinchuan Xu Affiliation: Microsoft Jialu Gao Affiliation: Microsoft Youkow Homma Affiliation: Microsoft Jian Jiao â thanks: Jian Jiao and Youkow Homma led the project. Affiliation: Microsoft Abstract Large language models (LLMs) are increasingly used to judge how well an advertisement matches a query, but the fairness of these judgments has received limited attention. We conduct a systematic study of fairness in relevance judgments made by LLMs for queries and advertisements. Our counterfactual framework examines the effects of advertiser identity and possible popularity, input language, and demographic wording. We study GPT-4o as a categorical relevance judge and a Qwen-7B model trained specifically for relevance prediction. The advertiser and language experiments use query and advertisement pairs sampled from real advertising logs. Controlled synthetic queries are used to study demographic associations in employment, housing, and credit. For both models, changing the advertiser identity or input language can alter the relevance assessment. Selected demographic comparisons also show patterns consistent with common stereotypes, particularly those involving gender and occupation. We further study mitigation during model inference and training. The results indicate that its effectiveness depends on whether advertiser information is relevant to the query and how advertiser labels are distributed in the training data. These findings can help advertising practitioners identify fairness risks and develop suitable mitigation methods for LLM relevance systems. Content Warning: This paper contains examples of potentially offensive content. 1 Introduction A relevance model estimates how well an advertisement matches a userâs query and intent. We study two model settings. GPT-4o assigns each query and advertisement pair a categorical label of good, fair, or bad. A Qwen-7B model trained for relevance prediction instead produces a relevance score. Ideally, these outputs should be determined by the semantic relationship between the query and the advertisement. In practice, they may also be influenced by demographic wording, input language, or the advertiserâs name. This sensitivity may affect both the advertisements shown to users and the way advertisers are treated. Fairness concerns in this setting involve both users and advertisers. Demographic terms in a query may activate stereotypes in consequential domains such as employment, housing, and credit. Users expressing the same intent in different languages may also receive different outcomes. Advertisements with the same content may be judged differently when they display different advertiser names, potentially favoring widely recognized brands over smaller advertisers. Such variation raises fairness concerns when information that does not alter the underlying relevance of the query and advertisement changes the modelâs judgment. Existing studies address related aspects of this problem, but largely separate brand fairness from ad-relevance modeling. Research on LLM-based recommender systems has documented popularity and brand-related preferences in movie and product recommendations, including differences between established and lesser-known brands (Lichtenberg et al., 2024; Chu and Hou, 2026). However, these studies examine LLM preferences in recommendation and brand-association tasks rather than the fairness of queryâad relevance judgments. Evaluating fairness in this task helps determine whether factors unrelated to the underlying queryâad match cause the model to treat users or advertisers differently. Such evidence can help advertising practitioners identify potential fairness risks and develop appropriate evaluation and mitigation strategies for LLM-based relevance systems. Separately, recent work has used fine-tuned language models for sponsored search relevance classification (Rokon et al., 2026) and LLM-generated judgments for advertiser keyphrase relevance labeling in large-scale advertising systems (Dey et al., 2025). These studies primarily focus on relevance accuracy, scalable labeling, and operational effectiveness rather than systematically evaluating fairness across users and advertisers. To our knowledge, this work provides one of the first systematic fairness evaluations of LLM-based queryâad relevance judgments. We apply a unified counterfactual framework to examine advertiser identity and potential popularity related effects, input language sensitivity, and demographic wording. The advertiser and language evaluations use queryâad pairs sampled from real advertising logs, while controlled synthetic queries are used to study demographic attributes that rarely appear explicitly in logged queries. We compare GPT-4o as a general-purpose LLM judge with a fine-tuned Qwen-7B relevance model, allowing us to examine fairness risks across different model types and output formulations. Our mitigation analysis evaluates company name masking during inference separately for queries that target a company and queries that do not. For confidentiality, advertiser names are anonymized for reporting only; the experimental inputs are unchanged. For the Qwen-7B relevance model, we also compare random downsampling of Retailer A examples with rebalancing based on the labels of Retailer A examples in the training data. Our experiments show that advertiser identity, input language, and demographic wording can affect relevance outputs even when the underlying query intent and advertisement content are kept the same. Both models produce different results across advertisers. In the selected comparisons, two widely recognized advertisers, Retailer A and Job Platform A, generally receive more favorable outcomes, although the results do not show a strict ranking based on advertiser popularity. Semantically equivalent inputs in different languages also lead to different outcomes in both models. Selected demographic comparisons show patterns consistent with common stereotypes, particularly those involving gender and occupation. Together, these findings reveal fairness risks in judgments made by a general language model and in predictions from a model trained specifically for relevance. In the mitigation experiments, removing company names has little effect on Qwen-7B performance when queries do not target a company, but lowers performance when they do. The GPT-4o experiment, which was conducted once, shows no observed loss in aggregate accuracy after company names are removed. In the training data experiment, rebalancing Retailer A examples according to their labels reduces Retailer Aâs relative advantage, while randomly removing Retailer A examples does not have the same effect. These findings support the use of query intent to determine when company names should be masked, as well as targeted changes to the distribution of labels in the training data. Our main contributions are as follows: ⢠We introduce a unified counterfactual framework for evaluating advertiser identity and popularity, input language, and demographic bias in LLM-based queryâad relevance judgments. The framework supports both categorical judgments from general-purpose LLMs and relevance probabilities from fine-tuned models. ⢠We conduct a multi perspective fairness evaluation of GPT-4o and fine-tuned Qwen-7B. The advertiser and language experiments are grounded in queryâad pairs sampled from real advertising logs, while controlled synthetic counterfactuals are used to examine demographic associations in employment, housing, and credit. ⢠We evaluate two practical mitigation strategies. The first masks company names during inference, and the second rebalances the training data. We examine masking separately for queries that target a company and queries that do not. We also compare random downsampling with rebalancing based on the training labels. 2 Related Work 2.1 Fairness in LLM-based recommendation Fairness in LLM-based recommendation can be considered from both the user and advertiser sides. From the user side, Sakib and Das (2024) focus on music, song, and book recommendations across diverse demographic and cultural groups. They construct queries related to different genders and regions and compare whether LLMs provide similar recommendations when demographic or cultural attributes vary. Instead of merely comparing the similarity of outputs, Deldjoo and Di Noia (2025) incorporate personal preferences into the fairness evaluation process. First, they identify the userâs personal preferences and ask the LLM to generate recommendations. They then add a sensitive attribute to the query and request recommendations again. Finally, they compare the two sets of recommendations to assess fairness. These studies motivate our user-side evaluation of demographic bias. However, our work focuses on queryâad relevance judgments rather than LLM-generated recommendation lists. From the item side, Jiang et al. (2024) examine popularity and genre bias. For items in different sensitive groups, the recommendation rate should align with the interaction rate of those items across all users. Zhao et al. (2025) consider both item-side and user-side bias. They define stereotype groups containing both users and items, and assess fairness by comparing the proportion of same-group items in recommendations with that in usersâ interaction histories. Our work examines potential advertiser popularity bias at the relevance assessment stage by changing the advertiser identity while keeping the query and all other ad content unchanged. Language bias can affect both users and advertisers. Users expressing the same intent in different languages may receive different outcomes, while advertisers presenting semantically equivalent ad content in different languages may also be evaluated differently. Previous multilingual evaluations have found substantial cross-language variation in LLM judgments (Fu and Liu, 2025). In the ad relevance setting, we examine this sensitivity using equivalent English, Chinese, and Finnish queryâad pairs and compare both categorical relevance judgments and predicted relevance probabilities. For GPT-4o, a duplicated English condition provides a reference for separating cross-language differences from the modelâs inherent output variability. More broadly, our work uses a unified counterfactual framework to study how advertiser identity, input language, and demographic wording affect relevance judgments made by LLMs for queries and advertisements. 2.2 Bias mitigation Many mitigation methods have been developed to address fairness issues in LLMs. For example, Liu et al. (2025) propose a fairness-aware LLM-based recommender system that filters out sensitive information. Jiang et al. (2024) apply a reweighting technique and introduce an unfairness penalty during post-training reranking to mitigate item-side bias. Dong et al. (2024) reduce LLM bias by tuning model parameters, incorporating a fairness loss, and providing fairness instructions in prompts. In our ad-relevance setting, we evaluate company-name masking at inference time separately for company-targeting and non-company-targeting queries. For the fine-tuned Qwen-7B model, we also compare random downsampling of Retailer A examples with label-aware rebalancing that changes the distribution of Retailer A training labels. 3 A Unified Framework for Bias Evaluation We develop a unified counterfactual framework to evaluate potential biases in both general purpose LLMs and task specific fine-tuned LLMs. Let =xii=1nD=\x_i\_i=1^n denote a collection of original inputs, where n is the number of inputs and each xix_i contains a queryâad pair. For a given bias attribute a, such as advertiser identity, demographic information, or input language, we construct a counterfactual transformation TaT_a that modifies attribute a while preserving all other information relevant to the labeling task. The resulting counterfactual input is xi(a)=Taâ(xi).x_i^(a)=T_a(x_i). The models considered in this study produce two different types of outputs. Let âłcatM_cat denote the collection of general purpose LLMs that produce categorical relevance judgments, and let âłprobM_prob denote the collection of task specific fine-tuned LLMs that produce relevance probabilities. For a general purpose LLM mââłcatm _cat, let Lmâ(x)â:=good,fair,badL_m(x) :=\good,fair,bad\ denote the categorical relevance judgment assigned to input x. Let Râ0,1Râ\0,1\ denote the binary relevance indicator, where R=1R=1 indicates that the queryâad pair is relevant and R=0R=0 indicates that it is irrelevant. For a task specific fine-tuned model mââłprobm _prob, let pmâ(x):=PrmâĄ(R=1âŁx)â[0,1]p_m(x):= _m(R=1 x)â[0,1] denote the predicted probability that the queryâad pair is relevant. Accordingly, we use output specific measures to quantify counterfactual sensitivity: changes in categorical judgments for general purpose LLMs and changes in predicted relevance probabilities for fine-tuned models. We investigate three potential sources of bias in ad-relevance assessment: advertiser identity and popularity, demographic attributes, and input language. The counterfactual construction is adapted to each type of bias. For advertiser bias, we compare each original query and advertisement pair with several versions in which the advertiser has been replaced. For demographic bias, we construct synthetic counterfactual pairs and paraphrases that refer to the same demographic group. For language bias, we compare semantically equivalent versions of the same query and advertisement pair in different languages. We then compare the model outputs using the measures defined for each type of output in the preceding section. Advertiser identity and popularity bias. We begin with queryâad pairs sampled from advertising logs. We restrict the analysis to queries that do not explicitly mention an advertiser, so that replacing the advertiser does not alter the semantic relationship between the query and the ad. For each original queryâad pair, we construct several counterfactual versions by replacing all occurrences of the original advertiser name in the ad title, description, and URL with a prespecified advertiser. The set of replacement advertisers is âŹ=Retailer A,Retailer B,Retailer C,Retailer D,Retailer E,Retailer F.B=\Retailer A,Retailer B,Retailer C,Retailer D,Retailer E,Retailer F\. Let xi(0)x_i^(0) denote the original queryâad pair and let bi(0)b_i^(0) denote its original advertiser. For each bââŹb , the corresponding counterfactual input is xi(b)=Tbi(0)âbâ(xi(0)),x_i^(b)=T_b_i^(0)â b (x_i^(0) ), where the transformation replaces the advertiser identity while leaving the query and the remaining ad content unchanged. All advertiser-specific versions are constructed from the same original queryâad pair, with the query and non-advertiser content held fixed. This matched design allows us to examine whether relevance judgments depend on advertiser identity and whether advertisers with different levels of market visibility receive systematically different outcomes. To illustrate the advertiser substitution procedure, we construct the following versions of the same queryâad pair: ⢠Original-advertiser version: black hat â Original Retailer Official Site â Save On black hat; ⢠Retailer A-substituted version: black hat â Retailer A Official Site â Save On black hat; ⢠Retailer C-substituted version: black hat â Retailer C Official Site â Save On black hat. The query and the non-advertiser content remain fixed across these versions, allowing us to assess how advertiser identity influences relevance judgments. Demographic bias. Because demographic attributes are rarely stated explicitly in observed search queries, we construct synthetic queryâad pairs. The synthetic pairs cover high-impact domains such as employment, housing, and credit (Meta, 2023). Within each template, we vary only the demographic attribute while keeping the requested service and ad content unchanged. The queries are synthetic, whereas the advertisements are real. The examples and Tables 6 and 7 display descriptive ad labels rather than the ad text used as model input. For example, a gender counterfactual pair is given by ⢠Original Query/Ad label: Jobs for female/ Engineering recruitment; ⢠Counterfactual Query/Ad label: Jobs for male/ Engineering recruitment. Since the advertised job is unchanged, the two inputs are expected to receive the same relevance judgment. A change in the model output therefore indicates sensitivity to the demographic attribute. To account for sensitivity to wording alone, we also construct a semantically equivalent query that refers to the same demographic group using a different expression: ⢠Within-group control Query/Ad label: Jobs for women/ Engineering recruitment. More generally, the synthetic templates combine queries concerning jobs, housing, or loans with ads representing occupations, housing options, or loan amounts. Each template is evaluated across multiple demographic groups and within-group paraphrases. Language bias. To evaluate language sensitivity, we translate each queryâad pair from English into non-English languages while preserving its semantic content. The English version and its translations form a set of multilingual counterfactual inputs. Under language invariant relevance assessment, semantically equivalent translations should receive consistent categorical judgments or similar predicted relevance probabilities. For example, we construct the following multilingual versions: ⢠English version: black hat â Retailer A Official Site â Save On black hat; ⢠Chinese version: éťč˛ĺ¸˝ĺâéśĺŽĺAĺŽćšç˝çŤâč´äš°éťč˛ĺ¸˝ĺĺŻäşŤäźć ; ⢠Finnish version: musta hattu â Vähittäiskauppiaan A virallinen sivusto â Sästä mustan hatun ostossa. We compare the English version with each non-English version. These comparisons allow us to examine whether the modelâs relevance assessment is stable across languages. 4 Experiments We evaluate whether ad-relevance judgments are sensitive to three attributes that should not, by themselves, change relevance: advertiser identity, input language, and demographic wording. We study both GPT-4o, which outputs one of three categorical labels (good, fair, or bad), and a Qwen-7B model fine-tuned on 700K queryâadâlabel examples, which outputs a probability of relevance (pRel). Following the counterfactual framework introduced in Section 3, we construct matched queryâad pairs that differ only in the attribute being evaluated, while preserving the query intent and all other ad content. Differences in model outputs across a matched pair therefore indicate sensitivity to the modified attribute. 4.1 Bias Evaluation 4.1.1 Advertiser identity and popularity bias Data and interventions. We select queries from real logs that do not mention a specific company and retain only those with at least four characters remaining after whitespace and punctuation are removed. We group the associated advertisements according to their content and focus on two categories: retail products and employment. The two evaluation sets contain 2,000 retail query and advertisement pairs and 878 employment query and advertisement pairs. For each retail advertisement, we replace every occurrence of the original advertiser name in the title, description, and URL with Retailer A, Retailer B, Retailer C, Retailer D, Retailer E, or Retailer F. For each employment advertisement, we replace the original advertiser name with either Job Platform A or Job Platform B. Within each set of counterfactual versions, the query and all advertisement content except the advertiser identity remain unchanged. Retailer A is a large and widely recognized retailer. Job Platform A is larger and more widely recognized than Job Platform B. The anonymized identifiers distinguish advertisers and do not indicate a ranking by size or popularity. GPT-4o. For GPT-4o, we summarize the categorical judgments using the bad decision rate (Bad DR) and the fair decision rate (Fair DR). Following the notation in Section 3, these metrics are defined as BadDRm=1nâi=1n[Lm(xi)=bad],BadDR_m= 1n _i=1^n1\! [L_m(x_i)=bad ], and FairDRm=1nâi=1n[Lm(xi)âbad,fair].FairDR_m= 1n _i=1^n1\! [L_m(x_i)â\bad,fair\ ]. Thus, Bad DR is the proportion of queryâad pairs labeled as bad, whereas Fair DR is the proportion labeled as either bad or fair. A lower value of either metric indicates more favorable relevance judgments. Because GPT-4o can assign different labels when the same input is evaluated multiple times, we run each advertiser condition ten times using the same 2,0002,000 query and advertisement pairs. Table 1 reports the decision rates averaged across the ten runs. To protect proprietary information, reported values are shifted by undisclosed constants. All statistical tests are computed using the original, unmodified data.These shifts preserve the ordering and absolute differences within each metric and do not affect the corresponding conclusions. Retailer A has the lowest Bad DR among the substituted retailers, at 14.6%, compared with 15.1%â16.6% for the other retailers. Its Fair DR is 59.4%, which is lower than those of Retailer C, Retailer D, Retailer E, and Retailer F, but slightly higher than Retailer Bâs 58.6%. The tests reject the null hypothesis for both metrics, indicating that Retailer A receives lower decision rates than the average of the other five substituted retailers. These results suggest that Retailer A, a large and widely recognized retailer, tends to receive more favorable GPT-4o judgments than most of the selected comparison retailers. However, because Retailer B has a lower Fair DR than Retailer A, the results do not imply a uniform ordering of retailers by popularity. Table 1: GPT-4o judgments after retailer substitution. Original Retailer A Retailer B Retailer C Retailer D Retailer E Retailer F p-value Bad DR (%) 11.7 14.6 15.1 15.8 15.8 16.6 15.7 <0.001<0.001 Fair DR (%) 51.0 59.4 58.6 62.4 61.5 61.1 60.6 <0.001<0.001 Results are averaged over ten runs of 2,000 matched queryâad pairs. For each metric, the reported two-sample t-test compares Retailer A with the average of Retailer B, Retailer C, Retailer D, Retailer E, and Retailer F across the repeated runs. To protect proprietary information, reported values are shifted by undisclosed constants. All statistical tests are computed using the original, unmodified data. Using the same experimental protocol, we evaluate Job Platform A and Job Platform B over ten independent repetitions on 878 queryâad pairs. For each metric, a one-sided two-sample t-test compares the ten measurements for Job Platform A with those for Job Platform B. The null hypothesis states that Job Platform Aâs mean decision rate is equal to or higher than Job Platform Bâs mean decision rate, whereas the alternative hypothesis states that Job Platform Aâs mean decision rate is lower. As shown in Table 2, Job Platform A has a lower Bad DR than Job Platform B (14.4% versus 15.3%) and a lower Fair DR (59.6% versus 64.5%). The tests reject the null hypothesis for both Bad DR and Fair DR , indicating that replacing the advertiser with Job Platform A produces more favorable GPT-4o judgments than replacing it with Job Platform B. Since Job Platform A is the larger and more widely recognized job-search platform in this comparison, the result is consistent with advertiser-popularity bias. Table 2: GPT-4o judgments after job search platform substitution. Original Job Platform A Job Platform B p-value Bad DR (%) 9.9 14.4 15.3 <0.001<0.001 Fair DR (%) 56.3 59.6 64.5 0.005 Results are averaged over ten runs of 878 matched queryâad pairs. For each metric, the reported two-sample t-test compares Job Platform A with Job Platform B across the repeated runs. To protect proprietary information, reported values are shifted by undisclosed constants. All statistical tests are computed using the original, unmodified data. Fine-tuned Qwen-7B. Unlike GPT-4o, the fine-tuned Qwen-7B model produces deterministic outputs under our inference configuration. Therefore, each queryâad pair is evaluated once, and no repeated experiments or t-tests are conducted. Table 3 reports the mean predicted relevance probability (pRel) from the fine-tuned Qwen-7B model. A higher pRel represents a more favorable relevance prediction. For retail ads, Retailer A obtains a higher pRel than Retailer B, Retailer C, Retailer D, and Retailer F, but a lower pRel than Retailer E. For job ads, Job Platform A obtains a higher pRel than Job Platform B (65.8% versus 63.1%). These patterns are broadly consistent with the GPT-4o results and suggest that prominent advertisers may receive more favorable predictions in some comparisons. Table 3: Fine-tuned Qwen-7B mean pRel after advertiser substitution. Retail-product ads Original Retailer A Retailer B Retailer C Retailer D Retailer E Retailer F pRel (%) 64.3 59.6 55.8 58.0 57.5 60.7 58.3 Job-related ads Original Job Platform A Job Platform B pRel (%) 64.5 65.8 63.1 The retail-product and job-related evaluations contain 2,000 and 878 matched queryâad pairs, respectively. To protect proprietary information, reported values are shifted by undisclosed constants. 4.1.2 Language bias We use 2,000 English queryâad pairs as the baseline and construct Chinese and Finnish counterparts through translation. To distinguish variation caused by language from the inherent variability of GPT-4o outputs, we include a second English condition that uses exactly the same inputs as the baseline. We evaluate the baseline, duplicated English, Chinese, and Finnish conditions over ten repetitions. For each repetition, consistency is defined as the percentage of the 2,000 queryâad pairs for which GPT-4o assigns the same categorical label in two conditions. For English, we compare the baseline with the duplicated English condition; for Chinese and Finnish, we compare each translated condition with the English baseline. Agreement between the two English conditions estimates the consistency expected when the language is unchanged, providing a reference against which the consistency of the translated inputs is assessed. Table 4 reports GPT-4oâs averaged consistency, Bad DR, and Fair DR across the three languages. For each metric, we use a t-test to compare the values obtained under the Chinese and Finnish conditions. The null hypothesis states that the mean metric value is the same for Chinese and Finnish inputs, whereas the alternative hypothesis states that the two means differ. Table 4: GPT-4o performance across input languages. Metric English Chinese Finnish ChineseâFinnish p value Consistency (%) 94.6 83.7 81.8 <0.001<0.001 Bad DR (%) 8.0 7.1 11.2 <0.001<0.001 Fair DR (%) 44.2 40.5 45.6 <0.001<0.001 Each language condition contains 2,000 semantically equivalent queryâad pairs. For each metric, the reported p-value is obtained from a t-test comparing the Chinese and Finnish conditions. To protect proprietary information, reported values are shifted by undisclosed constants. All statistical tests are computed using the original, unmodified data. GPT-4o achieves its highest consistency on English inputs (94.6%), compared with 83.7% for Chinese and 81.8% for Finnish. Chinese inputs receive the lowest Bad DR (7.1%) and Fair DR (40.5%), whereas Finnish inputs receive the highest Bad DR (11.2%) and Fair DR (45.6%). The Chinese Finnish differences are significant for all three metrics. These results suggest that GPT-4o produces more favorable relevance judgments for Chinese inputs than for Finnish inputs, although its labels are most consistent when the inputs remain in English. More generally, the language associated with greater consistency is not necessarily the language associated with more favorable labels. Unlike GPT-4o, the fine-tuned Qwen-7B model produces deterministic outputs under our inference configuration. We therefore evaluate each input once and report the mean predicted relevance probability without repeated experiments or t-tests. Table 5: Fine-tuned Qwen-7B mean pRel across input languages. Metric English Chinese Finnish pRel (%) 66.7 54.4 71.3 To protect proprietary information, reported values are shifted by undisclosed constants. The fine-tuned Qwen-7B model assigns the highest mean pRel to Finnish inputs (71.3%), followed by English inputs (66.7%), and the lowest mean pRel to Chinese inputs (54.4%). Thus, the model predicts greater relevance for the Finnish versions and lower relevance for the Chinese versions of the semantically equivalent queryâad pairs. Together, the GPT-4o and Qwen-7B results show that changing only the input language can lead to different relevance outcomes. 4.1.3 Demographic bias Because gender, race, and age are rarely explicit in logged queries, we use synthetic counterfactual queries in three consequential domains: employment, housing, and credit. Within each comparison, the ad is held fixed and only the demographic phrase in the query is changed. GPT-4o. Table 6 presents selected counterfactual cases with large differences in GPT-4oâs label distributions. Each queryâad pair is evaluated ten times, and the counts are reported as good:fair:bad. The employment results show a clear association between gender and occupation. Engineering jobs receive more favorable labels for queries referring to males (1:9:0) than for queries referring to females (0:3:7). In contrast, nurse jobs receive more favorable labels for queries referring to females (0:10:0) than for those referring to males (0:0:10). This pattern is consistent with common stereotypes that associate engineering with men and nursing with women. Queries referring to males also receive more favorable labels for both loan advertisements, suggesting an association between gender and credit products. Across the selected employment and housing cases, racial wording produces a consistent change in the label distribution. When the query wording changes from White people to Black people, the number of bad labels decreases from 8 to 5 for engineering jobs, from 8 to 2 for waiter/waitress jobs, from 8 to 0 for luxury housing, and from 6 to 0 for low cost housing. Most judgments move from bad to fair rather than to good. These results indicate that GPT-4oâs labeling behavior differs according to the racial wording in the query. The age results are more specific. GPT-4o associates young adults more strongly with waiter/waitress jobs and low cost housing than with middle aged or older adults. Table 6: GPT-4o predictions for demographic counterfactual pairs. Domain Ad label Group 1 G:F:B Group 2 G:F:B Panel A: Gender Employment Engineering recruitment Females 0:3:7 Males 1:9:0 Employment Registered-nurse recruitment Females 0:10:0 Males 0:0:10 Credit Personal loans up to $100K Females 0:10:0 Males 10:0:0 Credit Small personal loans Females 0:10:0 Males 8:2:0 Panel B: Race Employment Engineering recruitment White people 0:2:8 Black people 0:5:5 Employment Waiter/waitress recruitment White people 0:2:8 Black people 0:8:2 Housing Luxury house plans White people 0:2:8 Black people 0:10:0 Housing Low-cost houses White people 3:1:6 Black people 2:8:0 Panel C: Age Employment Engineering recruitment Young adults 0:6:4 Middle-aged/older adults 0:6:4 Employment Waiter/waitress recruitment Young adults 9:1:0 Middle-aged/older adults 0:10:0 Housing Luxury house plans Young adults 0:10:0 Middle-aged/older adults 0:10:0 Housing Low-cost houses Young adults 10:0:0 Middle-aged/older adults 0:9:1 Each queryâad pair is independently evaluated by GPT-4o ten times. G:F:B reports the numbers of good, fair, and bad labels across the ten repetitions. Within each row, the ad remains unchanged and only the demographic phrase in the query is modified. Ad labels are descriptive only; the models were evaluated on the corresponding real ad content. For confidentiality, the counts in this table have been perturbed and should not be interpreted as exact experimental counts. Fine-tuned Qwen-7B. Unlike GPT-4o, the fine-tuned Qwen-7B model is deterministic under our inference configuration, so each value in Table 7 represents a single pRel prediction. We also include alternative expressions for the same demographic groups, such as females and women, Black people and African Americans, and young adults and people aged 22 to 35, to examine sensitivity to wording. The employment results show a clear association between gender and occupation. For engineering jobs, queries referring to females receive a lower pRel than queries referring to males (0.270 versus 0.867). For nursing jobs, the pattern is reversed, with pRel values of 0.871 for females and 0.246 for males. The results for women, with values of 0.369 for engineering jobs and 0.862 for nursing jobs, suggest that this pattern is not limited to the word females. Patterns associated with race also appear. Compared with queries referring to White people, queries referring to Black people or African Americans generally receive lower pRel for engineering jobs and luxury housing, but higher pRel for waiter/waitress jobs and low cost housing. The findings for age are less stable. For example, engineering jobs receive a pRel of 0.373 for young adults and 0.893 for people aged 22 to 35, even though the two phrases describe similar age groups. Overall, the results suggest stereotypical associations involving gender and race. The age results also show substantial sensitivity to the choice of words used in the query. Table 7: Fine-tuned Qwen-7B pRel predictions for demographic queries. Domain Query group Ad label 1 pRel Ad label 2 pRel Panel A: Gender Employment Females Engineering recruitment 0.270 Nursing recruitment 0.871 Employment Males Engineering recruitment 0.867 Nursing recruitment 0.246 Employment Women Engineering recruitment 0.369 Nursing recruitment 0.862 Housing Females Luxury houses 0.764 Low-cost houses 0.633 Housing Males Luxury houses 0.746 Low-cost houses 0.511 Housing Women Luxury houses 0.709 Low-cost houses 0.618 Credit Females Loans up to $100K 0.930 Small loans 0.903 Credit Males Loans up to $100K 0.893 Small loans 0.854 Credit Women Loans up to $100K 0.912 Small loans 0.874 Panel B: Race Employment White people Engineering recruitment 0.581 Waiter/waitress recruitment 0.617 Employment Black people Engineering recruitment 0.398 Waiter/waitress recruitment 0.689 Employment African Americans Engineering recruitment 0.423 Waiter/waitress recruitment 0.603 Housing White people Luxury houses 0.625 Low-cost houses 0.611 Housing Black people Luxury houses 0.538 Low-cost houses 0.819 Housing African Americans Luxury houses 0.585 Low-cost houses 0.868 Credit White people Loans up to $100K 0.777 Small loans 0.622 Credit Black people Loans up to $100K 0.887 Small loans 0.877 Credit African Americans Loans up to $100K 0.843 Small loans 0.669 Panel C: Age Employment Young adults Engineering recruitment 0.373 Waiter/waitress recruitment 0.944 Employment Middle-aged/older adults Engineering recruitment 0.106 Waiter/waitress recruitment 0.658 Employment People aged 22 to 35 Engineering recruitment 0.893 Waiter/waitress recruitment 0.933 Housing Young adults Luxury houses 0.485 Low-cost houses 0.742 Housing Middle-aged/older adults Luxury houses 0.687 Low-cost houses 0.645 Housing People aged 22 to 35 Luxury houses 0.833 Low-cost houses 0.893 Credit Young adults Loans up to $100K 0.943 Small loans 0.904 Credit Middle-aged/older adults Loans up to $100K 0.935 Small loans 0.809 Credit People aged 22 to 35 Loans up to $100K 0.946 Small loans 0.929 Within each domain, the same two ads are evaluated using queries that differ only in their demographic wording. A higher pRel indicates greater predicted queryâad relevance. The fine-tuned Qwen-7B model is deterministic under our inference configuration, so each value represents a single prediction. Ad labels are descriptive only; the models were evaluated on the corresponding real ad content. To protect proprietary information, reported values are shifted by undisclosed constants. 4.2 Bias Mitigation We examine two approaches to mitigating advertiser popularity bias. At inference time, we hide company names in ads and evaluate the effect separately for company-targeting and non-company-targeting queries. At training time, we rebalance Retailer A examples in the 700K fine-tuning set by comparing random downsampling with the selective removal of good and fair examples. 4.2.1 Hiding company information at inference time We divide the test data labeled by humans into two subsets: 1,203 queryâad pairs in which the query does not target a specific company, and 519 pairs in which the query explicitly targets a company. For each pair, we remove the company name from the ad while leaving the query and all other ad content unchanged. For example, Book Now & Save Big at Travel Advertiser A! Always The Lowest Price Guarantee becomes Book Now & Save Big! Always The Lowest Price Guarantee. We then compare predictions based on the original and modified ads with the human labels. For GPT-4o, we evaluate each input once. Therefore, the results in Table 8 are intended as a demonstration rather than a statistical comparison across repeated runs. Hiding company information does not reduce the observed accuracy in either subset. Accuracy changes from 62.3% to 64.1% for queries that do not target a company and from 58.8% to 59.9% for queries that do. In both subsets, the Fair DR and Bad DR after masking move closer to the corresponding human-label rates. These descriptive results suggest that company information can be hidden without an observed loss in aggregate GPT-4o accuracy. For the fine-tuned Qwen-7B model, we report AUC on the same test data with human labels. AUC summarizes the modelâs predictive performance across different decision thresholds, with a higher value indicating better performance. For queries that do not target a company, hiding company information produces almost no change in AUC. For company-targeting queries, however, AUC decreases from 0.804 to 0.766. These results indicate that company information can be removed with little effect when it is unrelated to the query intent, but should be preserved when the user explicitly requests a particular company. Table 8: GPT-4o results before and after hiding company information. Query subset Metric Human labels Original GPT-4o GPT-4o without company No company mentioned Accuracy (%) â 62.3 64.1 Bad DR (%) 46.8 33.4 38.1 Fair DR (%) 83.3 75.3 75.8 Company mentioned Accuracy (%) â 58.8 59.9 Bad DR (%) 48.9 34.1 39.7 Fair DR (%) 83.8 74.2 80.0 To protect proprietary information, reported values are shifted by undisclosed constants. Table 9: Fine-tuned Qwen-7B test AUC before and after hiding company information. Query subset Original AUC Hidden-company AUC No company mentioned 0.814 0.814 Company mentioned 0.804 0.766 To protect proprietary information, reported values are shifted by undisclosed constants. 4.2.2 Reducing Retailer A examples during fine-tuning We next retrain Qwen-7B after downsampling Retailer A examples in the original 700K fine-tuning set. Table 10 compares the original model with two interventions using the same 2,000-pair advertiser-substitution evaluation. The first randomly removes half of all Retailer A examples. The second selectively removes half of Retailer Aâs good and fair examples. Random downsampling does not meaningfully change Retailer Aâs relative position: as in the baseline, Retailer A remains below Retailer E but above all other substituted retailers. Therefore, reducing the number of Retailer A examples without changing their label distribution does not appear to mitigate the modelâs relative preference for Retailer A. Selectively reducing Retailer Aâs good and fair examples produces a different pattern. Although Retailer Aâs pRel changes only slightly, from 59.6% to 59.7%, its relative ranking decreases. It falls below both Retailer E and Retailer F and becomes tied with Retailer C. This shift suggests that adjusting the label distribution of Retailer A examples can reduce the modelâs relative preference for the brand. The intervention does not completely remove differences among advertisers, but the results suggest that carefully rebalancing the training data is more effective than random downsampling in reducing bias related to advertiser popularity. Table 10: Fine-tuned Qwen-7B pRel (%) after Retailer A downsampling. Training data Original Retailer A Retailer B Retailer C Retailer D Retailer E Retailer F Original 700K 64.3 59.6 55.8 58.0 57.5 60.7 58.3 Remove half of all Retailer A ads 66.1 61.7 58.8 60.4 59.4 62.4 60.6 Remove half of Retailer A good/fair ads 65.4 59.7 58.1 59.7 59.0 61.7 60.1 To protect proprietary information, reported values are shifted by undisclosed constants. 5 Future work Future work could systematically examine how the proportions of advertiser examples and relevance labels in the training data affect the balance between fairness and predictive performance during model fine-tuning. Training models with different proportions of advertisers and labels may help identify data compositions that reduce disparities among advertisers while preserving relevance quality. Further research could also investigate how to mitigate language and demographic bias in LLM-based ad relevance tasks. Acknowledgments The authors used OpenAI models to assist with language editing of this manuscript. References Chu and Hou (2026) X. Chu and Y. Hou Incumbent advantage: brand bias and cognitive manipulation dynamics in llm recommendation systems. arXiv preprint arXiv:2606.17443. Cited by: §1. Deldjoo and Di Noia (2025) Y. Deldjoo and T. Di Noia Cfairllm: consumer fairness evaluation in large-language model recommender system. ACM Transactions on Intelligent Systems and Technology. Cited by: §2.1. Dey et al. (2025) S. Dey, H. Wu, and B. Li To judge or not to judge: using llm judgements for advertiser keyphrase relevance at ebay. arXiv preprint arXiv:2505.04209. Cited by: §1. Dong et al. (2024) X. Dong, Y. Wang, P. S. Yu, and J. Caverlee Disclosure and mitigation of gender bias in llms. arXiv preprint arXiv:2402.11190. Cited by: §2.2. Fu and Liu (2025) X. Fu and W. Liu How reliable is multilingual llm-as-a-judge?. In EMNLP (Findings), p. 11040â11053. Cited by: §2.1. Jiang et al. (2024) M. Jiang, K. Bao, J. Zhang, W. Wang, Z. Yang, F. Feng, and X. He Item-side fairness of large language model-based recommendation system. In Proceedings of the ACM Web Conference 2024, p. 4717â4726. Cited by: §2.1, §2.2. Lichtenberg et al. (2024) J. M. Lichtenberg, A. Buchholz, and P. SchwĂśbel Large language models as recommender systems: a study of popularity bias. arXiv preprint arXiv:2406.01285. Cited by: §1. Liu et al. (2025) W. Liu, B. Liu, J. Qin, X. Zhang, W. Huang, and Y. Wang Fairness identification of large language models in recommendation. Scientific Reports 15 (1), p. 5516. Cited by: §2.2. Meta (2023) Meta Toward fairness in personalized ads. Cited by: §3. Rokon et al. (2026) M. O. F. Rokon, A. Simion, W. Du, M. Wen, H. Yao, and K. Lee Enhancement of e-commerce sponsored search relevancy with llm. arXiv preprint arXiv:2607.03886. Cited by: §1. Sakib and Das (2024) S. K. Sakib and A. B. Das Challenging fairness: a comprehensive exploration of bias in llm-based recommendations. In 2024 IEEE International Conference on Big Data (BigData), p. 1585â1592. Cited by: §2.1. Zhao et al. (2025) Z. Zhao, W. Fan, Y. Wu, and Q. Li Investigating and mitigating stereotype-aware unfairness in llm-based recommendations. arXiv preprint arXiv:2504.04199. Cited by: §2.1.