The Anatomy of Socio-Demographics in Pluralistic Offensiveness Modeling: When and Where It Helps, and the Role of Reasoning

Yifan Wang; Leixin Zhang; Jinyi Mu; Yu Wang; Mayank Jobanputra; Isabel Valera; Vera Demberg

cs.AIOct 27, 2026

The Anatomy of Socio-Demographics in Pluralistic Offensiveness Modeling: When and Where It Helps, and the Role of Reasoning

Authors: Yifan Wang, Leixin Zhang, Jinyi Mu, Yu Wang, Mayank Jobanputra, Isabel Valera, Vera Demberg

Organizations: Saarland University · University of Tübingen · University of California San Diego · Utrecht University ·

Abstract

Subjective judgments such as offensiveness often vary substantially across individuals, motivating pluralistic modeling approaches that incorporate annotators' socio-demographic information to capture this variation. However, prior findings on the benefits of socio-demographic information for pluralistic modeling are mixed, calling for a systematic analysis of both its effectiveness and the mechanisms through which it improves prediction. In this work, we study $\textbf{when}$ and $\textbf{where}$ socio-demographics benefit pluralistic offensiveness modeling, as well as $\textbf{the role of explicit reasoning}$ in leveraging this information. Across three multi-annotator datasets with prompted and trained LLMs, we first show that demographic gains are small and inconsistent for individual predictions but become clearer at the group and population levels, because aggregation attenuates the idiosyncratic variation that masks small but systematic demographic signals. Second, using a reasoning and prediction format that explicitly decomposes individual judgment into an overall baseline and a demographic-specific deviation, we show that the gains arise from both components: demographic information improves baseline estimation and enables meaningful demographic-specific shifts, and interventions show that better component estimates improve final predictions. Finally, comparing direct demographic conditioning with explicit reasoning about how the text interacts with annotator demographics, we find that reasoning yields more accurate baseline and shift estimates as well as final predictions, whereas conditioning alone rarely produces informative shifts. These results suggest that the mixed evidence in prior work may partly stem from individual-level evaluation, and that the value of socio-demographic information depends on both how it is evaluated and how models are guided to use it.

Figures & tables

Figure 1: Example from D3CODE using the trained Qwen-4B model. Variants differ in whether they incorporate the annotator’s socio-demographic profile and the demographic-specific deviation , and socio-demographics-aware reasoning . Here, the model correctly estimates a non-offensive baseline and, given the historical sensitivity of the topic, a positive shift for the Japanese annotator, leading to the correct offensive individual prediction despite the non-offensive majority judgment.

Setting	Variant	Individual		Group				Population
Setting	Variant	F1 $\uparrow$	Acc $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho\uparrow$
Prompted	NoDemo	0.410	0.587	0.382	0.630	0.263	0.269	0.384	0.673	0.191	0.434
	DemoCond	0.414	0.589	0.390	0.636	0.230*	0.295*	0.389	0.680*	0.132*	0.526*
	DemoReason	0.426**	0.588	0.394*	0.639	0.235*	0.301*	0.394*	0.690**	0.128*	0.543**
Trained	NoDemo	0.357	0.583	0.375	0.631	0.198	0.291	0.412	0.684	0.120	0.473
	DemoCond	0.355	0.580	0.375	0.629	0.179*	0.303*	0.411	0.673	0.095*	0.512*

Table 1: Individual-, group- and population-level results for prompted and trained models. Scores are averaged across all dataset–model pairs. * / ** indicate significant improvement over NoDemo / over both NoDemo and the other demographic-aware variant (paired

t

-test,

p<0.05

). The benefits of incorporating socio-demographics become more apparent at higher levels of aggregation.

Setting	Variant	Baseline Estimation (Ann.)				Baseline Estimation (Item)
Setting	Variant	MSE $\downarrow$	MAE $\downarrow$	$\rho\uparrow$	W1 $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	$\rho\uparrow$	W1 $\downarrow$
Prompted	NoDemo	0.067	0.195	0.524	0.114	0.065	0.194	0.542	0.115
	DemoCond	0.061	0.186	0.529	0.102*	0.051*	0.172*	0.603*	0.101*
	DemoReason	0.059	0.183*	0.538	0.097*	0.050*	0.169*	0.601*	0.097
Trained	NoDemo	0.050	0.168	0.505	0.098	0.045	0.160	0.566	0.094
	DemoCond	0.047	0.164	0.510	0.090	0.038*	0.148*	0.604*	0.086

Table 2: Annotation- and item-level baseline estimation results for prompted and trained models. Scores are averaged across all dataset–model pairs. * / ** indicate significant improvement over NoDemo / over both NoDemo and the other demographic-aware variant (paired

t

-test,

p<0.05

). Avg. Individual Pred. averages individual predictions across annotators for each item and reports the best score across variant–setting pairs for each metric. Incorporating socio-demographic information, particularly through explicit reasoning, consistently improves baseline estimation.

Regime	Variant	Low disagreement				High disagreement
Regime	Variant	MSE	MAE	$\rho$	W1	MSE	MAE	$\rho$	W1
Prompted	DemoCond	+0.013	+0.019	+0.075	+0.013	+0.020	+0.041	+0.022	+0.042
Prompted	DemoReason	+0.014	+0.019	+0.069	+0.014	+0.026	+0.052	+0.017	+0.052
Trained	DemoCond	+0.003	+0.006	+0.034	+0.009	+0.012	+0.025	+0.024	+0.024
Trained	DemoReason	+0.005	+0.008	+0.048	+0.013	+0.014	+0.031	+0.091	+0.023

Table 3: Gain over NoDemo in item-level baseline estimation, stratified by disagreement (low/high disagreement: empirical baseline outside/inside

[0.4,0.6]

, 83.1%/16.9% of items). Positive values indicate improvement over NoDemo, with larger values indicating greater gains. Demographic-aware variants show larger gains on high-disagreement items.

Figure 2: Within-item standard deviation of baseline estimates for DemoCond and DemoReason across trained and prompted LLMs. DemoReason shows lower within-item variation than DemoCond for prompted models.

Figure 3: Model-emitted shift results for (a) global group tendency, shown as gain over NoDemo; (b) individual–majority deviations, shown as gain over random; and (c) item–group interactions, shown as sign agreement above chance across

|\gamma|

thresholds. Scores are averaged over dataset–model pairs; significance is assessed with paired

t

-tests (

p<0.05

). DemoReason consistently captures meaningful demographic effects across all three analyses.

Model	Variant	Individual	Group		Population
Model	Variant	F1 $\uparrow$	F1 $\uparrow$	MSE $\downarrow$	F1 $\uparrow$	MSE $\downarrow$
Llama-8B (Trained)	DemoCond	0.453/0.540/0.364	0.442/0.589/0.424	0.190/0.105/0.170	0.549/0.667/0.444	0.094/0.084/0.106
Llama-8B (Trained)	DemoReason	0.474/0.600/0.470	0.450/0.522/0.439	0.188/0.145/0.196	0.526/0.600/0.541	0.059/0.065/0.068
Llama-70B (Prompted)	DemoCond	0.438/0.554/0.312	0.425/0.544/0.257	0.178/0.140/0.200	0.435/0.652/0.387	0.095/0.068/0.077
Llama-70B (Prompted)	DemoReason	0.458/0.557/0.401	0.443/0.550/0.379	0.180/0.166/0.201	0.500/0.630/0.465	0.084/0.068/0.085

Table 4: Performance under joint interventions on crowd and shift estimates. Each cell reports Original / Oracle / Adversarial performance. Bold marks cells following the expected ordering Oracle

>

Original

>

Adversarial. Better baseline and shift estimates are not only associated with, but also lead to, improved final predictions.

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Test Annots. / Items	# Annots. / Item	Axes (# groups)
D3CODE	2,507 / 77	32.6	region (8), country (21), age (3), gender (3), SES (10)
DICES	10,849 / 134	81.0	gender (4), race (6), age (3), education (3)
POPQUORN	1,308 / 150	8.7	gender (3), race (6), age (10), education (5)

Appendix

Table 5: Dataset statistics and socio-demographic coverage.

Model	Variant	Individual		Group				Population
Model	Variant	F1 $\uparrow$	Acc $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho\uparrow$
Llama-3B	NoDemo	0.455	0.583	0.376	0.623	0.377	0.247	0.324	0.645	0.302	0.401
	DemoCond	0.453	0.584	0.380	0.624	0.320	0.271	0.334	0.639	0.194	0.533
	DemoReason	0.425	0.569	0.360	0.608	0.280	0.227	0.337	0.663	0.134	0.498
Qwen-4B	NoDemo	0.426	0.593	0.391	0.638	0.260	0.282	0.384	0.682	0.189	0.443
	DemoCond	0.364	0.575	0.362	0.618	0.211	0.270	0.360	0.664	0.115	0.494

Appendix

Table 6: Performance for all prompted models and baselines across different levels. Metric scores are averaged over three datasets, and Overall averages across backbones. * denotes significant improvement over NoDemo; ** denotes significant improvement over both NoDemo and the other variant across model–dataset pairs (paired

t

-test,

p<0.05

Model	Variant	Individual		Group				Population
Model	Variant	F1 $\uparrow$	Acc $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho\uparrow$
Llama-3B	NoDemo	0.342	0.576	0.339	0.610	0.204	0.261	0.340	0.634	0.124	0.424
	DemoCond	0.308	0.562	0.312	0.592	0.188	0.239	0.310	0.600	0.097	0.439
	DemoReason	0.334	0.570	0.356	0.611	0.188	0.282	0.378	0.657	0.073	0.497
Qwen-4B	NoDemo	0.381	0.587	0.404	0.645	0.197	0.317	0.467	0.720	0.117	0.512
	DemoCond	0.372	0.589	0.402	0.642	0.171	0.331	0.444	0.686	0.094	0.540

Appendix

Table 7: Performance for all trained models and baselines across different levels. Metric scores are averaged over three datasets, and Overall averages across backbones. * denotes significant improvement over NoDemo; ** denotes significant improvement over both NoDemo and the other variant across model–dataset pairs (paired

t

-test,

p<0.05

Variant	Individual		Group				Population
Variant	F1 $\uparrow$	Acc $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho$ $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho$ $\uparrow$
NoDemo	0.362	0.572	0.363	0.602	0.222	0.247	0.391	0.642	0.150	0.387
DemoCond	0.340	0.566	0.354	0.604	0.199	0.259	0.381	0.642	0.108	0.466
DemoReason	0.345	0.561	0.361	0.600	0.198	0.261	0.394	0.659	0.088	0.484

Appendix

Table 8: Cross-dataset transfer performance of trained models. For each source dataset and model, performance is averaged over the two target datasets as transfer targets; each cell then reports the mean across all source-dataset–model pairs.

Variant	Split	Individual		Group				Population
Variant	Split	F1 $\uparrow$	Acc $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho$ $\uparrow$	F1 $\uparrow$	Acc $\uparrow$	MSE $\downarrow$	$\rho$ $\uparrow$
DemoReason	ID	0.349	0.546	0.327	0.570	0.255	0.179	0.408	0.632	0.070	0.482
DemoReason	OOD	0.352	0.556	0.342	0.572	0.310	0.157	0.397	0.628	0.122	0.368

Appendix

Table 9: Unseen demographic-group transfer results of Llama-8B DemoReason on D3CODE. We report both ID results, where demographic groups are observed during training, and OOD results, where the evaluated groups are held out and unseen during training.

Figure 4: NoDemo example from a trained Qwen-9B model on D3CODE. Without access to the annotator’s socio-demographic profile, NoDemo substantially underestimates the crowd baseline (0.1 vs. 0.4306) and incorrectly predicts the judgment as non-offensive.

Figure 5: DemoCond example from a trained Qwen-9B model on D3CODE. Compared with NoDemo, DemoCond recognizes greater disagreement in how annotators may perceive the post and therefore provides a more accurate crowd baseline estimate (0.3 vs. 0.4306). However, without explicitly reasoning about the perception of the specific annotator, DemoCond predicts shift =0, fails to capture the individual–majority deviation, and still makes an incorrect final prediction.

Figure 6: DemoReason example from a trained Qwen-9B model on D3CODE. Compared with DemoCond, DemoReason provides a more accurate crowd estimate (0.4 vs. 0.4306). With explicit socio-demographic reasoning, DemoReason identifies that a Canadian annotator may be more sensitive to a post targeting Canadians and therefore predicts shift =+1, correctly capturing the individual–majority deviation and yielding the correct offensive final prediction.

Model	Variant	Baseline Estimation (Ann.)				Baseline Estimation (Item)
Model	Variant	MSE $\downarrow$	MAE $\downarrow$	$\rho\uparrow$	W1 $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	$\rho\uparrow$	W1 $\downarrow$
Llama-3B	NoDemo	0.060	0.186	0.320	0.098	0.059	0.185	0.333	0.097
	DemoCond	0.060	0.193	0.286	0.098	0.041	0.161	0.501	0.112
	DemoReason	0.068	0.204	0.236	0.078	0.044	0.168	0.394	0.098
Qwen-4B	NoDemo	0.093	0.238	0.555	0.155	0.089	0.234	0.565	0.154
	DemoCond	0.083	0.223	0.516	0.129	0.066	0.198	0.588	0.113

Appendix

Table 10: Annotation- and item-level baseline estimation results for all prompted models and baselines. Metric scores are averaged over three datasets, and Overall averages across backbones. * denotes significant improvement over NoDemo; ** denotes significant improvement over both NoDemo and the other variant across model–dataset pairs (paired

t

-test,

p<0.05

Model	Variant	Baseline Estimation (Ann.)				Baseline Estimation (Item)
Model	Variant	MSE $\downarrow$	MAE $\downarrow$	$\rho\uparrow$	W1 $\downarrow$	MSE $\downarrow$	MAE $\downarrow$	$\rho\uparrow$	W1 $\downarrow$
Llama-3B	NoDemo	0.063	0.198	0.424	0.135	0.058	0.189	0.485	0.133
	DemoCond	0.054	0.184	0.452	0.116	0.044	0.171	0.545	0.115
	DemoReason	0.054	0.180	0.440	0.089	0.041	0.158	0.565	0.093
Qwen-4B	NoDemo	0.039	0.144	0.594	0.055	0.034	0.137	0.653	0.046
	DemoCond	0.041	0.151	0.559	0.056	0.033	0.135	0.640	0.051

Appendix

Table 11: Annotation- and item-level baseline estimation results for all trained models and baselines. Metric scores are averaged over three datasets, and Overall averages across backbones. * denotes significant improvement over NoDemo; ** denotes significant improvement over both NoDemo and the other variant across model–dataset pairs (paired

t

-test,

p<0.05

Model	Variant	Individual	Group		Population
Model	Variant	F1 $\uparrow$	F1 $\uparrow$	MSE $\downarrow$	F1 $\uparrow$	MSE $\downarrow$
Llama-8B (Trained)	DemoCond	0.453/0.441/0.337	0.442/0.477/0.410	0.190/0.152/0.168	0.549/0.650/0.457	0.094/0.081/0.096
Llama-8B (Trained)	DemoReason	0.474/0.478/0.443	0.450/0.453/0.421	0.188/0.185/0.197	0.526/0.595/0.556	0.059/0.059/0.063
Llama-70B (Prompted)	DemoCond	0.438/0.510/0.335	0.425/0.511/0.270	0.178/0.164/0.195	0.435/0.667/0.353	0.095/0.075/0.078
Llama-70B (Prompted)	DemoReason	0.458/0.549/0.432	0.443/0.534/0.413	0.180/0.186/0.196	0.500/0.632/0.468	0.084/0.077/0.073

Appendix

Table 12: Performance under baseline-only interventions on baseline and shift estimates for Llama-8B (trained) and Llama-70B (prompted). Each cell reports Original / Oracle / Adversarial performance. Bold marks cells following the expected ordering Oracle

>

Original

>

Adversarial.

Model	Variant	Individual	Group		Population
Model	Variant	F1 $\uparrow$	F1 $\uparrow$	MSE $\downarrow$	F1 $\uparrow$	MSE $\downarrow$
Llama-8B (Trained)	DemoCond	0.453/0.412/0.333	0.442/0.483/0.398	0.190/0.131/0.165	0.549/0.378/0.432	0.094/0.110/0.108
Llama-8B (Trained)	DemoReason	0.474/0.582/0.492	0.450/0.527/0.450	0.188/0.149/0.200	0.526/0.500/0.600	0.059/0.075/0.071
Llama-70B (Prompted)	DemoCond	0.438/0.538/0.499	0.425/0.495/0.463	0.178/0.167/0.183	0.435/0.480/0.519	0.095/0.095/0.088
Llama-70B (Prompted)	DemoReason	0.458/0.525/0.500	0.443/0.495/0.482	0.180/0.175/0.188	0.500/0.528/0.538	0.084/0.086/0.084

Appendix

Table 13: Performance under shift-only interventions on baseline and shift estimates for Llama-8B (trained) and Llama-70B (prompted). Each cell reports Original / Oracle / Adversarial performance. Bold marks cells following the expected ordering Oracle

>

Original

>

Adversarial.

User Modeling

User modeling represents people's interests, knowledge, goals, and changing behavior to personalize predictions and AI assistance.

See all 154 papers

LLM-as-a-Demographic: Whom Sociodemographic Prompting Helps, and Whom It Hurts
Aug 31, 2026Daniela Occhipinti, Andrea Piergentili, Marco GueriniLLM-as-a-Judge Intersectional Bias in Language Models
Fondazione Bruno Kessler, Via Sommarive 18, Povo, Trento, Italy · Almawave Labs, Via Di Casal Boccone, 188/190, Rome
Demographic Pluralism: Inference-Time Modeling of Pluralistic Human Preference Distributions
Sep 29, 2026Meng-Chen Wu, Qipin Chen, Ansh Jain +3Human Preference Modeling Inference-Time Language Model Alignment
Amazon Science
When Does Demographic Information Help? Data and Modeling Regimes for Perspective-Aware Hate Speech Detection
May 26, 2026Weibin Cai, Reza ZafaraniHate Speech Detection
Data Lab, EECS Department Syracuse University

Aug 31, 2026J/K move · Enter open · S save

LLM-as-a-Demographic: Whom Sociodemographic Prompting Helps, and Whom It Hurts

Daniela Occhipinti, Andrea Piergentili, Marco Guerini

Fondazione Bruno Kessler, Via Sommarive 18, Povo, Trento, Italy · Almawave Labs, Via Di Casal Boccone, 188/190, Rome

LLM-as-a-Judge Intersectional Bias in Language Models LLM Prompting Social Bias in Language Models Language Model Bias Evaluation Bias in LLM-as-a-Judge+1

Open paper AskPDF arXiv

Large language models (LLMs) are increasingly used as judges for subjective tasks, where annotators disagree and the relevant question is not only how accurate a judge is, but whose judgments it reproduces. Sociodemographic prompting conditions the judge on an annotator's demographic profile to align its judgments with the corresponding group's. We test whether this alignment emerges distributionally, comparing the predicted label distributions of 23 open-weight LLMs on three subjective tasks against those of real annotator groups, under three conditions: no demographic information, single-attribute profiles, and intersectional profiles over gender, age, race, and education. Three findings emerge. First, a judge prompted with no demographics is not perspective-neutral: models best reproduce the judgments of White, college-educated annotators. Second, demographic conditioning is asymmetric: it moves the judge toward majority groups and away from minority groups, most strongly on offensiveness, where intersectional profiles amplify the harm. Third, by comparing base and instruct models we identify instruction-tuning as a possible source of the asymmetry. Demographic conditioning should therefore be used with caution to estimate group judgments: conditioning moves predictions away from the reference distributions of the minority groups the method is often invoked to serve.

Abstract

User Modeling

Explore similar work

LLM-as-a-Demographic: Whom Sociodemographic Prompting Helps, and Whom It Hurts

Demographic Pluralism: Inference-Time Modeling of Pluralistic Human Preference Distributions

When Does Demographic Information Help? Data and Modeling Regimes for Perspective-Aware Hate Speech Detection