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

Meng-Chen Wu; Qipin Chen; Ansh Jain; Tess Wood; Zhe Du; Si-Chi Chin

cs.AISep 29, 2026

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

Authors: Meng-Chen Wu, Qipin Chen, Ansh Jain, Tess Wood, Zhe Du, Si-Chi Chin

Organizations: Amazon Science

Abstract

Large language models (LLMs) are increasingly used in culturally sensitive settings, where alignment requires representing diverse preferences within populations. Yet existing methods model populations at coarse demographic or community levels and overlook within-group variation. We introduce Demographic Pluralism, an inference-time framework that estimates population-level opinion distributions without opinion-distribution training data or task-specific fine-tuning by generating multiple perspectives within demographically grounded groups. Across four backbones on GlobalOpinionQA and VITAL, it reduces Jensen-Shannon distance by 8.4%-26.4% over Modular Pluralism. Among weighted, equal-weighted, and inverse-weighted aggregation, equal weighting performs best overall; group-level error also increases with group weight, helping explain weighted aggregation's weaker performance.

Figures & tables

Figure 1: Overview of the Demographic Pluralism framework. (1) Demographic Group Generation selects question-relevant attributes and constructs plausible demographically supported groups. (2) Diverse Perspectives Generation elicits multiple viewpoints for each group to capture within-group variation. (3) Opinion Distribution Generation obtains an answer distribution from each perspective and aggregates the distributions within and across groups. The framework operates at inference time without observed survey responses or fine-tuning.

Baseline
Method	Mistral-7B-Instruct		Phi-3.5-Instruct		Qwen2.5-7B-Instruct		Qwen3-14B
Method	Global $\downarrow$	VITAL $\downarrow$	Global $\downarrow$	VITAL $\downarrow$	Global $\downarrow$	VITAL $\downarrow$	Global $\downarrow$	VITAL $\downarrow$
Direct Prompting	.3224 ( $5\text{e-}5$ )	.3645 ( $1\text{e-}4$ )	.4178 ( $8\text{e-}5$ )	.3925 ( $5\text{e-}4$ )	.3906 ( $2\text{e-}5$ )	.3576 ( $6\text{e-}4$ )	.4076 ( $8\text{e-}6$ )	.3412 ( $6\text{e-}6$ )
CoT	.3905 ( $1\text{e-}3$ )	.3862 ( $4\text{e-}3$ )	.4806 ( $1\text{e-}3$ )	.4639 ( $6\text{e-}3$ )	.3893 ( $6\text{e-}4$ )	.3846 ( $3\text{e-}3$ )	.4234 ( $4\text{e-}4$ )	.4025 ( $4\text{e-}3$ )
Modular Pluralism	.3292 ( $7\text{e-}4$ )	.3574 ( $8\text{e-}4$ )	.3366 ( $6\text{e-}4$ )	.3226 ( $2\text{e-}3$ )	.3551 ( $5\text{e-}4$ )	.3462 ( $2\text{e-}3$ )	.3369 ( $1\text{e-}4$ )	.3044 ( $9\text{e-}4$ )
Demographic Pluralism

Table 1: Jensen–Shannon distance (JSD) across four language models on GlobalOpinionQA and its VITAL subset (mean over 3 independent runs; standard deviation in parentheses). Bold marks the best method per model–evaluation-set pair, and underlining marks the second-best. Lower is better. Improvement is the relative JSD reduction of the best Demographic Pluralism method over Modular Pluralism.

Panel A: Full question-disjoint test (15,893 instances)
Method	Uses training-set opinion distributions?	LLM fine-tuned?	Mistral-7B	Phi-3.5	Qwen2.5-7B	Qwen3-14B
Direct Prompting	—		.327	.417	.394	.406
CoT	—		.393	.480	.389	.423
Modular Pluralism	—		.305	.316	.308	.315
Cao et al.			.217	.362	.539	.314
Demographic Pluralism	—		.292	.272	.270	.281

Table 2: Question-disjoint comparison by use of opinion-distribution training data. Panel A reports the full test set; Panel B reports the P2P-compatible test set. Values are JSD (

\downarrow

); bold marks the best result for each backbone. A black check mark and dash indicate that training-set opinion distributions are used and unused, respectively. Snowflake and fire symbols indicate that the LLM is frozen and fine-tuned, respectively. P2P trains only ensemble weights, whereas Modular Pluralism fine-tunes on external community data. We report equal-weighted Demographic Pluralism , the strongest full-corpus aggregation strategy (see Table 1 ).

Figure 2: Alignment on pluralistic (gold dominance

\leq 0.60

) and dominant (

>0.60

) cases. Demographic Pluralism achieves the lowest JSD and TailJSD in both regimes, with the largest gaps on pluralistic cases where baselines collapse onto a single option. Lower is better; Top-2 mass is reported in Appendix H .

Figure 3: JSD (

\downarrow

) by resource tier on GlobalOpinionQA (left) and VITAL (right). Demographic Pluralism is lowest in every tier.

Method	Country	Subgroup	Pearson	Spearman
	JSD $\downarrow$	JSD $\downarrow$
Demographic Pluralism	0.2106	0.2671	0.799	0.779
w/o Multiple Perspectives	0.3139	0.3811	0.695	0.699

Table 3: Country- and subgroup-level alignment on the microdata-covered GlobalOpinionQA subset using Qwen2.5-7B-Instruct and equal-weighted aggregation. The full method improves over the ablation at both levels and yields a stronger country–subgroup correlation. Lower JSD is better.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Statistic	GlobalOpinionQA	VITAL
Number of country–question instances	28,763	1,676
Initial demographic combinations	2,790,345	168,691
After zero-probability filtering	1,784,375 (63.9%)	106,996 (63.4%)
After plausibility filtering (LLM)	918,078 (32.9%)	55,322 (32.8%)
Mean demographic groups per instance	31.9	33.0
Median demographic groups per instance	16	17

Appendix

Table 4: Statistics of demographic group construction and perspective generation per country–question instance. Percentages are relative to the initial enumeration of demographic combinations.

Figure 4: Alignment performance (JSD,

\downarrow

) as a function of the number of retained perspectives (

K

) on GlobalOpinionQA (left) and VITAL (right). Solid lines show Demographic Pluralism under four language models; dashed horizontal lines indicate the corresponding Modular Pluralism baseline. Performance stabilizes around

K=20

–

30

, with as few as

K=5

perspectives already outperforming Modular Pluralism on most models.

Demographic Pluralism (Full)
Method	Mistral-7B-Instruct		Phi-3.5-Instruct		Qwen2.5-7B-Instruct		Qwen3-14B
Method	Global $\downarrow$	VITAL $\downarrow$	Global $\downarrow$	VITAL $\downarrow$	Global $\downarrow$	VITAL $\downarrow$	Global $\downarrow$	VITAL $\downarrow$
weighted	.2884 ( $2\text{e-}4$ )	.3275 ( $6\text{e-}4$ )	.2749 ( $4\text{e-}4$ )	.2843 ( $2\text{e-}3$ )	.2750 ( $2\text{e-}4$ )	.2588 ( $1\text{e-}3$ )	.2835 ( $3\text{e-}4$ )	.2461 ( $1\text{e-}3$ )
equal-weighted	.2869 ( $1\text{e-}4$ )	.3276 ( $5\text{e-}4$ )	.2669 ( $4\text{e-}4$ )	.2808 ( $5\text{e-}4$ )	.2681 ( $2\text{e-}4$ )	.2547 ( $5\text{e-}4$ )	.2757 ( $2\text{e-}4$ )	.2386 ( $4\text{e-}4$ )
inverse-weighted	.2870 ( $2\text{e-}4$ )	.3275 ( $6\text{e-}4$ )	.2670 ( $4\text{e-}4$ )	.2815 ( $2\text{e-}3$ )	.2682 ( $2\text{e-}4$ )	.2553 ( $1\text{e-}3$ )	.2758 ( $2\text{e-}4$ )	.2394 ( $2\text{e-}3$ )
w/o demographic groups	.3015 ( $6\text{e-}4$ )	.3245 ( $4\text{e-}3$ )	.3114 ( $8\text{e-}4$ )	.3203 ( $5\text{e-}2$ )	.3058 ( $4\text{e-}4$ )	.2988 ( $4\text{e-}3$ )	.3147 ( $3\text{e-}4$ )	.2829 ( $3\text{e-}3$ )

Appendix

Table 5: Ablation analysis of Demographic Pluralism. We report Jensen–Shannon Distance (JSD) across four language models and two evaluation sets (mean over 3 independent runs; standard deviation in parentheses). Bold indicates the best and underline the second-best per model–benchmark pair. Lower is better.

Attribute schema beyond age+gender	Gain on Global	Gain on VITAL
$+$ ethnicity	$+0.010$	$+0.121$
$+$ religion $+$ ethnicity	$+0.023$	$+0.199$
$+$ education $+$ religion $+$ ethnicity	$+0.024$	$+0.200$

Appendix

Table 6: Average JSD improvement over age and gender alone on the schema-complete matched subset (Qwen2.5-7B-Instruct, equal-weighted aggregation). This separate fixed-schema diagnostic is not directly comparable to the whole-benchmark ablation in Table 5 .

Dataset	Cutoff	Pluralistic	Dominant
GlobalOpinionQA	0.55	13,311 (46.3%)	15,452 (53.7%)
	0.60	16,334 (56.8%)	12,429 (43.2%)
	0.65	18,984 (66.0%)	0 9,779 (34.0%)
VITAL	0.55	0 511 (30.5%)	1,165 (69.5%)
	0.60	0 700 (41.8%)	0 976 (58.2%)
	0.65	0 877 (52.3%)	0 799 (47.7%)

Appendix

Table 7: Proportion of pluralistic and dominant cases under different gold-dominance thresholds. Pluralistic cases are defined as question–country pairs where no response option exceeds the specified cutoff.

		Pluralistic			Dominant
Cutoff	Method	JSD $\downarrow$	Top-2 $\uparrow$	TailJSD $\downarrow$	JSD $\downarrow$	Top-2 $\uparrow$	TailJSD $\downarrow$
0.55	Direct Prompting	0.30	0.53	0.28	0.34	0.72	0.24
	Modular Pluralism	0.32	0.51	0.30	0.34	0.70	0.26
	Demographic Pluralism	0.26	0.56	0.25	0.31	0.75	0.21
0.60	Direct Prompting	0.30	0.55	0.27	0.35	0.74	0.24
	Modular Pluralism	0.32	0.53	0.30	0.34	0.72	0.26

Appendix

Table 8: Robustness analysis on GlobalOpinionQA under different gold-dominance thresholds. Demographic Pluralism consistently improves pluralistic alignment while preserving dominant-case behavior across all cutoffs.

		Pluralistic			Dominant
Cutoff	Method	JSD $\downarrow$	Top-2 $\uparrow$	TailJSD $\downarrow$	JSD $\downarrow$	Top-2 $\uparrow$	TailJSD $\downarrow$
0.55	Direct Prompting	0.32	0.59	0.30	0.38	0.73	0.31
	Modular Pluralism	0.30	0.57	0.30	0.38	0.68	0.27
	Demographic Pluralism	0.27	0.62	0.27	0.35	0.75	0.24
0.60	Direct Prompting	0.32	0.62	0.30	0.39	0.73	0.31
	Modular Pluralism	0.31	0.59	0.29	0.39	0.69	0.27

Appendix

Table 9: Robustness analysis on VITAL under different gold-dominance thresholds. Demographic Pluralism consistently improves pluralistic alignment while maintaining dominant-case behavior.

Case Study (Dataset: GlobalOpinionQA; Country: Pakistan)

Question: Protecting the environment should be given priority, even if it causes slower economic growth and some loss of jobs.

Options: A. Completely agree B. Mostly agree C. Mostly disagree D. Completely disagree

Modular Pluralism (full set of fixed perspectives).

news_left:

Text: As a Pakistani citizen, I strongly believe that the statement is very important and is of utmost importance. My preference will be that we should do both—protect the environment and at the same time, continue to have a strong and healthy economy with job creation. While it is necessary to have economic growth, our growth must be sustainable. We need to take steps which are responsible, where we do not create jobs that require us to damage the environment. However, if it means sacrificing short term economic benefits for long term environmental gains, we must be willing to do so. A strong and healthy economy can be achieved without having to come at the cost of the environment. Therefore, my answer will be “B. Mostly agree.”

Distribution: [0.04, 0.95, 0.00, 0.01]

Appendix

Figure 5: Qualitative case study (part 1 of 3): modular pluralism perspectives.

Demographic Pluralism (demographically grounded perspectives; Perspectives 1–9).

Perspective 1 (15–64Y male, Muslim):

Background & Perspective: Muhammad is a 25-year-old male from Karachi, Pakistan. He is a devout Muslim and attends mosque regularly. He works as a salesman at a local market and earns a moderate income. He is married with two children and lives in a small apartment in the city. He enjoys watching cricket and spending time with his family. Protecting the environment is crucial for the future of our children and the country. Economic growth and job creation should not come at the expense of our planet’s well-being.

Distribution: [0.33, 0.55, 0.07, 0.04]

Perspective 2 (15–64Y male, Muslim):

Background & Perspective: Muhammad is a 25-year-old male from Karachi, Pakistan. He is a devout Muslim and attends mosque regularly. He works as a salesman at a local market and earns a moderate income. He is married with two children and lives in a small apartment in the city. While I understand the importance of protecting the environment, I also believe that economic growth and job creation are necessary for the prosperity of our country. A balance must be struck between the two.

Distribution: [0.05, 0.23, 0.47, 0.25]

Appendix

Figure 6: Qualitative case study (part 2 of 3): Demographic Pluralism perspectives (1–9).

Demographic Pluralism (demographically grounded perspectives; continued).

Perspective 10 (65+Y female, Muslim):

Background & Perspective: Umm Zainab recognizes the importance of economic growth and job creation, but she also understands the long-term consequences of neglecting the environment. She believes that while economic growth is important, it should not come at the expense of the environment. She sees the two as interconnected and believes that protecting the environment will ultimately lead to a more prosperous and sustainable future for all.

Distribution: [0.34, 0.55, 0.06, 0.05]

Perspective 11 (65+Y female, Muslim):

Background & Perspective: Umm Zainab is a pragmatist and believes that protecting the environment should be given priority, but only if it can be done in a way that does not harm the economy or cause unnecessary hardship for people. She recognizes the challenges of balancing environmental protection with economic growth and believes that a middle ground can be found that benefits both.

Distribution: [0.09, 0.37, 0.30, 0.24]

Appendix

Figure 7: Qualitative case study (part 3 of 3): Demographic Pluralism perspectives (10–11) and final aggregated predictions.

Dataset	Method	High-resource	Medium-resource	Low-resource
		(39.6% / 32.6%)	(42.9% / 43.7%)	(17.6% / 23.6%)
GlobalOpinionQA	Direct Prompting	0.3249	0.3145	0.3367
	Modular Pluralism	0.3379	0.3204	0.3268
	Demographic Pluralism	0.2851	0.2840	0.2982
VITAL	Direct Prompting	0.3528	0.3655	0.3786
	Modular Pluralism	0.3536	0.3509	0.3720

Appendix

Table 10: Average Jensen–Shannon Distance (JSD) by country resource level. Percentages indicate the proportion of samples in each resource tier ( GlobalOpinionQA / VITAL ). Lower values indicate better alignment.

Method	Dataset	F-statistic	p-value	Significance
Direct Prompting	GlobalOpinionQA	37.94	$<10^{-6}$	Significant
Direct Prompting	VITAL	4.25	0.014	Significant
Modular Pluralism	GlobalOpinionQA	38.02	$<10^{-6}$	Significant
Modular Pluralism	VITAL	2.78	0.062	Not Significant
Demographic Pluralism	GlobalOpinionQA	19.15	$<10^{-6}$	Significant
Demographic Pluralism	VITAL	31.56	$<10^{-6}$	Significant

Appendix

Table 11: One-way ANOVA results testing whether alignment performance (JSD) differs across high-, medium-, and low-resource country groups. Significance is evaluated at

\alpha=0.05

. While resource level has a statistically significant effect for most method–dataset pairs, Modular Pluralism does not exhibit significant variation across resource tiers on the VITAL subset. Importantly, statistical significance reflects detectability of differences rather than their magnitude.

Method	Intermediate generator	Global JSD $\downarrow$	VITAL JSD $\downarrow$
CoT	Mistral-7B	0.3893	0.3846
CoT	Qwen3-32B	0.4788	0.4504
Demographic Pluralism	Mistral-7B	0.2681	0.2547
Demographic Pluralism	Qwen3-32B	0.2788	0.3095

Appendix

Table 12: Alignment (JSD

\downarrow

) with Qwen2.5-7B-Instruct as the downstream opinion model. Replacing the Mistral-7B intermediate generator with Qwen3-32B does not improve Demographic Pluralism ; each configuration remains substantially better than CoT using the same intermediate generator.

Metric	Global	VITAL
Cross-model perspective sim.	$0.500\pm 0.115$	$0.540\pm 0.111$
Cross-model background sim.	$0.594\pm 0.144$	$0.599\pm 0.137$
Within-Mistral perspective sim.	$0.767\pm 0.116$	$0.788\pm 0.109$
Within-Qwen perspective sim.	$0.575\pm 0.121$	$0.587\pm 0.123$
Mistral avg. perspectives/group	2.73	2.81
Qwen3-32B avg. perspectives/group	2.99	2.99

Appendix

Table 13: Semantic similarity of generated artifacts between Mistral-7B and Qwen3-32B (sentence embeddings, cosine similarity), across 665,264 matched (question, demographic group) pairs. Cross-model similarity is moderate (

\sim

0.50), and Qwen3-32B produces more numerous and internally diverse perspectives, yet does not improve downstream alignment.

Benchmark	Method	Artifact (s)	Opinion (s)	Total (s)
GlobalOpinionQA ( $n$ =28,763)	Direct Prompting	0	854	854
	CoT	73,101	461	73,562
	Modular Pluralism	48,402	11,454	59,856
	Demographic Pluralism	45,578	16,306	61,884
VITAL ( $n$ =1,676)	Direct Prompting	0	92	92
	CoT	4,771	71	4,842

Appendix

Table 14: Wall-clock time (seconds) decomposed into artifact generation and opinion generation. All methods use Mistral-7B-Instruct on a single A100 GPU.

Benchmark	Resource Tier	N (groups)	Spearman $\rho$	$p$ -value	JSD (bot. 1/3)	JSD (top 1/3)
GlobalOpinionQA	High	436,145	$-0.006$	$1.7\times 10^{-4}$	0.1300	0.1323
	Medium	359,202	$+0.010$	$1.0\times 10^{-9}$	0.1329	0.1369
	Low	122,191	$+0.059$	$1.3\times 10^{-95}$	0.1138	0.1295
VITAL	High	22,449	$+0.065$	$1.4\times 10^{-22}$	0.0965	0.1131
	Medium	23,698	$+0.137$	$1.0\times 10^{-99}$	0.0922	0.1111
	Low	9,151	$+0.085$	$4.4\times 10^{-16}$	0.1063	0.1216

Appendix

Table 15: Spearman correlation (

\rho

) between per-group demographic weight and per-group prediction error (JSD) on GlobalOpinionQA and VITAL (Qwen2.5-7B-Instruct), stratified by country resource level. Positive

\rho

indicates that high-weight demographic groups produce less accurate predictions. The rightmost columns report mean JSD for the lowest- and highest-weight terciles.

	Mistral-7B-Instruct						Phi-3.5-Instruct						Qwen2.5-7B-Instruct						Qwen3-14B
	Global			VITAL			Global			VITAL			Global			VITAL			Global			VITAL
Method	JSD	EMD	Hyb	JSD	EMD	Hyb	JSD	EMD	Hyb	JSD	EMD	Hyb	JSD	EMD	Hyb	JSD	EMD	Hyb	JSD	EMD	Hyb	JSD	EMD	Hyb
Direct Prompting	.3224	.2535	.2661	.3645	.2847	.2988	.4178	.2967	.3116	.3925	.2768	.3116	.3906	.2871	.3004	.3576	.2380	.2732	.4076	.3033	.3153	.3412	.2312	.2521
CoT	.3905	.2905	.3118	.3862	.2754	.3039	.4806	.3307	.3558	.4639	.3131	.3426	.3893	.2857	.2980	.3846	.2670	.2876	.4234	.3008	.3160	.4025	.2761	.3006
Modular Pluralism	.3292	.2489	.2670	.3574	.2857	.3065	.3366	.2341	.2512	.3226	.2445	.2537	.3551	.2371	.2598	.3462	.2180	.2636	.3369	.2328	.2522	.3044	.2106	.2368
Ours (WEIGHTED)	.2884	.2372	.2437	.3275	.2750	.2835	.2749	.2061	.2127	.2843	.2178	.2285	.2750	.2113	.2106	.2588	.1930	.2062	.2835	.2130	.2131	.2461	.1867	.1981

Appendix

Table 16: Alignment under JSD, EMD, and the question-aware Hybrid metric (all

\downarrow

) in the full-corpus protocol. Bold marks the best result, and underlining marks the second-best result within each backbone, evaluation set, and metric.

Panel A: 131-country question-disjoint test (15,893 instances, 1,191 questions)
Method	Mistral-7B	Phi-3.5	Qwen2.5-7B	Qwen3-14B
Direct Prompting	.256	.271	.252	.257
CoT	.295	.314	.249	.262
Modular Pluralism	.213	.203	.194	.196
Cao et al.	.149	.267	.361	.238
Demographic Pluralism	.211	.181	.169	.169

Appendix

Table 17: Normalized EMD (

\downarrow

) on question-disjoint GlobalOpinionQA test sets. Panel A covers all 131 test countries; Panel B uses the 19-country P2P-compatible subset. Bold marks the best result within each backbone and panel.

Panel A: 131-country question-disjoint test (15,893 instances, 1,191 questions)
Method	Mistral-7B	Phi-3.5	Qwen2.5-7B	Qwen3-14B
Direct Prompting	.285	.323	.313	.317
CoT	.334	.378	.307	.329
Modular Pluralism	.253	.247	.244	.247
Cao et al.	.180	.305	.437	.271
Demographic Pluralism	.247	.218	.210	.215

Appendix

Table 18: Question-aware Hybrid distance (

\downarrow

) on the same question-disjoint test sets. The metric uses EMD for ordinal questions and JSD for nominal questions. Bold marks the best result within each backbone and panel.

Explore similar work

PLURAL: A Global Dataset for Value Alignment
Jul 9, 2026Dhruv Agarwal, Anya Shukla, Tanya Goyal +1Preference Datasets Pluralistic Alignment
Cornell University
Population Fidelity: Evaluating Population Representativeness in LLMs
Sep 28, 2026Neemias B. da Silva, Martin Lukk, Ali Sutani +5Large Language Model Bias Demographics
University of Toronto, Canada · Federal University of Technology – Paran´a, Brazil
PALMs: Using Multi Construct-Grounded Rationales for Modeling Population Preferences in LLMs
Aug 2, 2026Priyanka Dey, Brihi Joshi, Preyashi Poddar +2Large Language Model Alignment Social Reasoning
University of Southern California · Information Sciences Institute

Jul 9, 2026, cs.CLJ/K move · Enter open · S save

PLURAL: A Global Dataset for Value Alignment

Dhruv Agarwal, Anya Shukla, Tanya Goyal, Aditya Vashistha

Cornell University

Preference Datasets Pluralistic Alignment Natural Language Datasets Value Alignment Value Diversity+1

Open paper AskPDF arXiv

Large language models (LLMs) are used worldwide, yet disproportionately reflect Western values, limiting their ability to represent diverse value systems. We introduce PLURAL, a large-scale, value-focused preference dataset grounded in the Integrated Values Survey (IVS), a nationally representative survey spanning 92 countries. Using a two-stage generation pipeline, we transform survey responses into synthetic preference triplets that preserve normative value signals while producing realistic scenarios. We release an initial version of PLURAL containing ~500,000 preference triplets representing people in 20 diverse countries. We evaluate PLURAL in three ways: (i) dataset-level validation showing that it preserves both cross-country value differences and within-country diversity from the original survey; (ii) automated evaluation showing that training on PLURAL improves alignment with target countries' cultural profiles, reducing mean absolute error by up to 27.7% relative to strong baselines; and (iii) blind human evaluation with 176 evaluators in India, Brazil, and Japan, who judge PLURAL-aligned responses as more representative of their national values. Together, these results show that PLURAL contains learnable signal for value steering, offering a scalable resource for pluralistic alignment. Dataset: https://huggingface.co/datasets/agdhruv/plural-alignment

Abstract

Explore similar work

PLURAL: A Global Dataset for Value Alignment

Population Fidelity: Evaluating Population Representativeness in LLMs

PALMs: Using Multi Construct-Grounded Rationales for Modeling Population Preferences in LLMs