Parametric Social Identity Injection and Diversification in Public Opinion Simulation

Hexi Wang; Yujia Zhou; Bangde Du; Qingyao Ai; Yiqun Liu

cs.CLMar 17, 2026

Parametric Social Identity Injection and Diversification in Public Opinion Simulation

Authors: Hexi Wang, Yujia Zhou, Bangde Du, Qingyao Ai, Yiqun Liu

Organizations: DCST, Tsinghua University Beijing, China · Quancheng Laboratory Jinan, Shandong, China

Abstract

Large language models (LLMs) have recently been adopted as synthetic agents for public opinion simulation, offering a promising alternative to costly and slow human surveys. Despite their scalability, current LLM-based simulation methods fail to capture social diversity, producing flattened inter-group differences and overly homogeneous responses across demographic groups. We identify this limitation as a Diversity Collapse phenomenon in LLM hidden representations, where distinct social identities become increasingly indistinguishable across layers. Motivated by this observation, we propose Parametric Social Identity Injection (PSII), a general framework that injects explicit, parametric representations of demographic attributes and value orientations directly into intermediate hidden states of LLMs. Unlike prompt-based persona conditioning, PSII enables fine-grained and controllable identity modulation at the representation level. Extensive experiments on the World Values Survey using multiple open-source LLMs show that PSII significantly improves distributional fidelity and diversity, reducing KL divergence to real-world survey data while enhancing overall diversity. This work provides new insights into representation-level control of LLM agents and advances scalable, diversity-aware public opinion simulation.

Figures & tables

Figure 1. Layer-wise scatter plots of final-token hidden states for 500 simulated agents (top) and an illustration of Diversity Collapse in Transformer hidden states (bottom). In the top panels, red points denote baseline methods and gray points denote PSII-generated agents; the reported scores measure the average spatial dispersion of representations in each layer. The bottom panel depicts the Diversity Collapse phenomenon. merge

Figure 2. Overview of the Parametric Social Identity Injection (PSII) mechanism. From left to right: Identity Construction , including agent profile construction and identity vector construction; Parametric Injection , including noise addition and hierarchical injection; Performance Evaluation of the simulated agents. Illustration of the PSII framework showing the pipeline from Identity Construction (agent profile and identity vector creation), through Parametric Injection (adding noise and hierarchical injection into hidden states), to Performance Evaluation of the simulated agents.

Model	Method	Beliefs & Life		Social Integration		Political Engagement		Economic Progress		Overall
		KL $\downarrow$	ED $\downarrow$	KL $\downarrow$	ED $\downarrow$	KL $\downarrow$	ED $\downarrow$	KL $\downarrow$	ED $\downarrow$	KL $\downarrow$	ED $\downarrow$
Qwen2.5-7B	Direct	1.5182	0.6664	0.8575	0.7406	2.2222	0.7625	1.6283	0.7932	1.3915	0.7340
	High-Temp	1.1941	0.4754	0.7074	0.5981	1.9653	0.6266	1.1072	0.5272	1.1605	0.5778
	Multilingual	1.0946	0.3706	0.5547	0.5112	1.9297	0.5000	1.1030	0.5371	1.0568	0.4811
	DivReq	1.5442	0.6671	0.8383	0.7464	2.2166	0.7587	1.5282	0.7283	1.3813	0.7329
	PE	1.4812	0.4050	0.7182	0.6292	1.8852	0.4847	1.5095	0.6046	1.2209	0.5443

Table 1. Main experimental results on the WVS dataset. We report KL divergence and Entropy Deviation (ED) for each method across four question categories and overall. Best-performing results are highlighted in bold.

Figure 3. Response distributions for a randomly selected question from each of the four categories. PSII more closely matches the empirical response diversity observed in human survey data, while baseline methods often concentrate on a few options. Response distributions for a randomly selected question from each of the four categories. \textcolor{myred}{PSII more closely matches the empirical response diversity observed in human survey data}, while baseline methods often concentrate on a few options.

Setting	Qwen2.5-7B		Qwen2.5-14B		Llama-3.1-8B		Mistral-24B
	KL ↓	ED ↓	KL ↓	ED ↓	KL ↓	ED ↓	KL ↓	ED ↓
PSII (full model)	0.4843	0.0319	0.5814	0.2123	0.4017	0.0040	0.5607	0.0774
PSII w/o value vector	0.5264	0.0348	0.6574	0.3040	0.4638	0.0416	0.6254	0.1723
PSII w/o demographic vectors	0.9705	0.3859	0.7893	0.3878	0.6681	0.2729	0.7185	0.3312
PSII w/o prompt-based profile	0.5024	0.0151	0.6688	0.2527	0.5575	0.0630	0.6149	0.0474
PSII w/o parametric noise	0.7237	0.2261	0.7332	0.3294	0.5673	0.2068	0.6813	0.2808

Table 2. Ablation study results on PSII across different models. Each row shows the impact of removing one component on accuracy and diversity metrics.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Original WVS Category	Question IDs
Social Values, Attitudes & Stereotypes	Q1–Q45
Happiness and Well-Being	Q46–Q56
Social Capital, Trust & Organizational Membership	Q57–Q105
Economic Values	Q106–Q111
Corruption	Q112–Q120
Migration	Q121–Q130

Appendix

Table 3. Original value categories and question mappings in the World Values Survey (WVS).

Question ID	Question Topic
Q260	Respondent’s sex
Q263	Native-born or immigrant status
Q271	Living with parents or parents-in-law
Q273	Marital status
Q275	Highest completed education level
Q279	Employment status and working hours

Appendix

Table 4. Demographic attributes used for constructing demographic vectors.

Figure 4. Language distribution in the WVS dataset. Language distribution in the WVS dataset.

Setting	Claude	DeepSeek	GPT-4o	4o-V1	4o-V2	4o-V3	4o-V4	4o-S1	4o-S123	4o-S42	GPT-5-mini
Avg. similarity	0.8742	0.9337	0.9341	0.9353	0.9378	0.9196	0.9304	0.9357	0.9387	0.9397	0.9043

Appendix

Table 5. Average pairwise semantic similarity of demographic descriptions within each vector-construction setting. Similarity is computed using paraphrase-multilingual-MiniLM-L12-v2.

Setting	KL $\downarrow$	ED $\downarrow$
GPT-4o, default setting	0.4843	0.0319
GPT-4o, seed = 1	0.5221	0.0294
GPT-4o, seed = 123	0.5142	0.0245
GPT-4o, seed = 42	0.4884	0.0173
GPT-5-mini, default setting	0.4997	0.0148
Average	0.5017	0.0236

Appendix

Table 6. Sensitivity analysis of demographic-vector construction on Qwen2.5-7B. We reconstruct demographic vectors under different instruction-model and seed settings and report downstream KL divergence and Entropy Deviation (ED).

Sample group	KL $\downarrow$	ED $\downarrow$
Group A: original sample	0.4843	0.0319
Group B: resample 1	0.4742	0.0302
Group C: resample 2	0.4660	0.0231
Group D: resample 3	0.4362	0.0272
Group E: resample 4	0.5051	0.0324
Average	0.4732	0.0290

Appendix

Table 7. Sampling robustness analysis on Qwen2.5-7B. We independently sample five groups of 100 respondents from the WVS dataset and report the overall KL divergence and Entropy Deviation (ED).

Model	Method	Beliefs & Life		Social Integration		Political Engagement		Economic Progress		Overall
		JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$
Qwen2.5-7B	Direct	0.6822	0.2264	0.5625	0.3157	0.7013	0.3067	0.7611	0.1671	0.6330	0.2884
	High-Temp	0.6167	0.2012	0.5116	0.2836	0.6356	0.2731	0.6460	0.1370	0.5722	0.2574
	Multilingual	0.5350	0.1747	0.4487	0.2488	0.5681	0.2341	0.6421	0.1373	0.5070	0.2247
	DivReq	0.7007	0.2375	0.5632	0.3167	0.6908	0.3005	0.7494	0.1649	0.6337	0.2893
	PE	0.5909	0.1799	0.5068	0.2861	0.5463	0.2224	0.7050	0.1524	0.5435	0.2414

Appendix

Table 8. Main experimental results on the WVS dataset. We report JS divergence and MAE for each method across four question categories and overall. Best-performing results are highlighted in bold.

Setting	Qwen2.5-7B		Qwen2.5-14B		Llama-3.1-8B		Mistral-24B
Setting	JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$	JS $\downarrow$	MAE $\downarrow$
PSII (full model)	0.2650	0.1277	0.3491	0.1573	0.2593	0.1303	0.2971	0.1419
PSII w/o value vector	0.2975	0.1373	0.3998	0.1768	0.2807	0.1424	0.3451	0.1603
PSII w/o demographic vectors	0.4526	0.2033	0.4368	0.1954	0.3946	0.1739	0.4157	0.1862
PSII w/o prompt-based profile	0.2795	0.1363	0.3678	0.1664	0.3102	0.1617	0.3069	0.1494
PSII w/o parametric noise	0.3753	0.1778	0.4171	0.1846	0.3478	0.1599	0.3904	0.1753

Appendix

Table 9. Ablation study results on PSII across different models. Each row shows the impact of removing one component on JS divergence and MAE. Removing any module results in performance degradation.

Figure 5. The effects of injecting demographic attributes into different network layers. It illustrates how layer selection impacts simulation accuracy (KL divergence) and diversity (normalized entropy). Visualization showing which demographic features are injected at low, intermediate, and upper layers, and their impact on KL divergence and normalized entropy.

Layer Setting	KL $\downarrow$	ED $\downarrow$
L1: Optimal Layer Configuration (OLC)	0.4843	0.0319
L2: Random Layer Selection (1)	1.0467	0.0812
L3: Random Layer Selection (2)	1.4093	0.2621
L4: Global Layer Shift (+2 from OLC)	0.5496	0.1046
L5: Global Layer Shift (–2 from OLC)	0.4858	0.0852
L6: Single-Layer Injection (50% depth)	0.9798	0.0367

Appendix

Table 10. Layer-wise injection sensitivity analysis on Qwen2.5-7B.

Figure 6. Layer-wise scatter plots of final-token hidden states for 500 simulated agents in Llama-3.1-8B . Red points correspond to baseline methods, while gray points correspond to agents generated using PSII. The reported scores measure the average spatial dispersion of representations in each layer. Layer-wise scatter plot of final-token hidden states for 500 simulated agents. Red points correspond to baseline methods, while gray points correspond to agents generated using PSII.

Figure 7. Layer-wise scatter plots of final-token hidden states for 500 simulated agents in Qwen2.5-7B . Red points correspond to baseline methods, while gray points correspond to agents generated using PSII. The reported scores measure the average spatial dispersion of representations in each layer. Layer-wise scatter plot of final-token hidden states for 500 simulated agents. Red points correspond to baseline methods, while gray points correspond to agents generated using PSII.

Figure 8. Layer-wise scatter plots of final-token hidden states for 500 simulated agents in Qwen2.5-14B . Red points correspond to baseline methods, while gray points correspond to agents generated using PSII. The reported scores measure the average spatial dispersion of representations in each layer. Layer-wise scatter plot of final-token hidden states for 500 simulated agents. Red points correspond to baseline methods, while gray points correspond to agents generated using PSII.

Figure 9. Layer-wise scatter plots of final-token hidden states for 500 simulated agents in Mistral-24B . Red points correspond to baseline methods, while gray points correspond to agents generated using PSII. The reported scores measure the average spatial dispersion of representations in each layer. Layer-wise scatter plot of final-token hidden states for 500 simulated agents. Red points correspond to baseline methods, while gray points correspond to agents generated using PSII.

Large Language Model-Based Social Simulation

Large language model-based social simulation studies how language-model agents represent people and interact to produce social behavior.

See all 255 papers

Apr 8, 2026

Restoring Heterogeneity in LLM-based Social Simulation: An Audience Segmentation Approach

Large Language Models (LLMs) are increasingly used to simulate social attitudes and behaviors, offering scalable "silicon samples" that can approximate human data. However, current simulation practice often collapses diversity into an "average persona," masking subgroup variation that is central to social reality. This study introduces audience segmentation as a systematic approach for restoring heterogeneity in LLM-based social simulation. Using U.S. climate-opinion survey data, we compare six segmentation configurations across two open-weight LLMs (Llama 3.1-70B and Mixtral 8x22B), varying segmentation identifier granularity, parsimony, and selection logic (theory-driven, data-driven, and instrument-based). We evaluate simulation performance with a three-dimensional evaluation framework covering distributional, structural, and predictive fidelity. Results show that increasing identifier granularity does not produce consistent improvement: moderate enrichment can improve performance, but further expansion does not reliably help and can worsen structural and predictive fidelity. Across parsimony comparisons, compact configurations often match or outperform more comprehensive alternatives, especially in structural and predictive fidelity, while distributional fidelity remains metric dependent. Identifier selection logic determines which fidelity dimension benefits most: instrument-based selection best preserves distributional shape, whereas data-driven selection best recovers between-group structure and identifier-outcome associations. Overall, no single configuration dominates all dimensions, and performance gains in one dimension can coincide with losses in another. These findings position audience segmentation as a core methodological approach for valid LLM-based social simulation and highlight the need for heterogeneity-aware evaluation and variance-preserving modeling strategies.

Xiaoyou Qin, Zhihong Li, Xiaoxiao Cheng

School of Journalism, Fudan University · College of Media and International Culture, Zhejiang University

Large Language Model-Based Social Simulation User Modeling LLM Evaluation

Ask PDF

Mar 28, 2026

Persona-Based Simulation of Human Opinion at Population Scale

What does it mean to model a person, not merely to predict isolated responses, preferences, or behaviors, but to simulate how an individual interprets events, forms opinions, makes judgments, and acts consistently across contexts? This question matters because social science requires not only observing and predicting human outcomes, but also simulating interventions and their consequences. Although large language models (LLMs) can generate human-like answers, most existing approaches remain predictive, relying on demographic correlations rather than representations of individuals themselves. We introduce SPIRIT (Semi-structured Persona Inference and Reasoning for Individualized Trajectories), a framework designed explicitly for simulation rather than prediction. SPIRIT infers psychologically grounded, semi-structured personas from public social media posts, integrating structured attributes (e.g., personality traits and world beliefs) with unstructured narrative text reflecting values and lived experience. These personas prompt LLM-based agents to act as specific individuals when answering survey questions or responding to events. Using the Ipsos KnowledgePanel, a nationally representative probability sample of U.S. adults, we show that SPIRIT-conditioned simulations recover self-reported responses more faithfully than demographic persona and reproduce human-like heterogeneity in response patterns. We further demonstrate that persona banks can function as virtual respondent panels for studying both stable attitudes and time-sensitive public opinion.

Mao Li, Frederick G. Conrad

Institute for Social Research, University of Michigan, 426 Thompson St., Ann Arbor, 48104, MI, US.

Human Behavior Simulation Large Language Model-Based Social Simulation Persona-Conditioned Generation Computational Social Science Opinion Dynamics

Ask PDF

Aug 20, 2026

Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories

Large language models (LLMs) offer a scalable approach to social simulation, but their credibility depends on how agents are constructed. Existing methods can partially reproduce population-level patterns, yet often fail to capture human-like diversity. Our analysis shows that static-profile agents exhibit stronger demographic separation and within-group compression than humans, a pattern consistent with identity essentialism: demographic labels can encourage models to treat group-average tendencies as individual traits, homogenizing responses within groups. We argue that this limitation arises from two related factors: sparse, static agent representations and the limited ability of prompt-only memory to persistently integrate experience. Inspired by complementary memory systems, we propose LifeMem, a longitudinal memory framework that combines structured life-event retrieval with agent-specific parametric memory for experience integration. Experiments on Understanding Society with three LLMs show that LifeMem improves alignment with human data in terms of response distributions, overall and within-group diversity, and patterns of within-person response change across life stages. These findings highlight the value of longitudinal life-event memory for constructing more faithful and dynamically evolving social agents.

Hexi Wang, Yujia Zhou, Bangde Du +7

Department of Computer Science and Technology, Tsinghua University · Quan Cheng Laboratory

Human Behavior Simulation LLM Agent Memory Large Language Model-Based Social Simulation Demographic Bias in Language Models Memory-Augmented Language Model Agents

Ask PDF

Abstract

Large Language Model-Based Social Simulation

Explore similar work

Restoring Heterogeneity in LLM-based Social Simulation: An Audience Segmentation Approach

Persona-Based Simulation of Human Opinion at Population Scale

Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories