PolicySim: An LLM-Based Agent Social Simulation Sandbox for Proactive Policy Optimization

Renhong Huang; Ning Tang; Jiarong Xu; Yuxuan Cao; Qingqian Tu; Sheng Guo; Bo Zheng; Huiyuan Liu

cs.SIMar 20, 2026

PolicySim: An LLM-Based Agent Social Simulation Sandbox for Proactive Policy Optimization

Authors: Renhong Huang, Ning Tang, Jiarong Xu, Yuxuan Cao, Qingqian Tu, Sheng Guo, Bo Zheng, Huiyuan Liu,

Organizations: Zhejiang University, Hangzhou, China · Fudan University, Shanghai, China · HKUST, Hong Kong, China · University of Nottingham, Nottingham, United Kingdom ·

Abstract

Social platforms serve as central hubs for information exchange, where user behaviors and platform interventions jointly shape opinions. However, intervention policies like recommendation and content filtering, can unintentionally amplify echo chambers and polarization, posing significant societal risks. Proactively evaluating the impact of such policies is therefore crucial. Existing approaches primarily rely on reactive online A/B testing, where risks are identified only after deployment, making risk identification delayed and costly. LLM-based social simulations offer a promising pre-deployment alternative, but current methods fall short in realistically modeling platform interventions and incorporating feedback from the platform. Bridging these gaps is essential for building actionable frameworks to assess and optimize platform policies. To this end, we propose PolicySim, an LLM-based social simulation sandbox for the proactive assessment and optimization of intervention policies. PolicySim models the bidirectional dynamics between user behavior and platform interventions through two key components: (1) a user agent module refined via supervised fine-tuning (SFT) and direct preference optimization (DPO) to achieve platform-specific behavioral realism; and (2) an adaptive intervention module that employs a contextual bandit with message passing to capture dynamic network structures. Experiments show that PolicySim can accurately simulate platform ecosystems at both micro and macro levels and support effective intervention policy.

Figures & tables

Figure 1. Compared to A/B testing, which reactively assesses interventions by deploying and learning only after outcomes are observed, PolicySim proactively assesses and optimizes intervention policies prior to deployment via feedback.

	Scale	Relation	IP.	AI.	Env.
PolicySim	1000	✓	✓	✓	X & Weibo
Oasis ( Yang et al., 2024 )	1M	✓	✓	✗	X & Reddit
Agent4rec ( Zhang et al., 2024 )	1000	✓	✓	✗	Movie Rec
HiSim ( Mou et al., 2024a )	700	✗	✗	✗	X
Stopia ( Zhou et al., 2025 )	2	✗	✗	✗	-

Table 1. Comparison of our system with recent social simulation frameworks. Scale: number of LLM agents; Relation: whether follow/follower links evolve; IP/AI: presence of intervention policy and adaptive interventions; Env.: underlying platform (“–” means unspecified).

Figure 2. The architecture of PolicySim is composed of two main modules: the User Agent Module and the Intervention Policy Module. The User Agent Module contains detailed components including user profiles, memory, user relations, and behavioral models. The bottom-left panel illustrates how we train agent tailored for social media environments to capture realistic user behaviors. The Intervention Policy Module instantiates typical platform mechanisms such as recommender systems and exposure control. By simulating intervention policies within sandbox framework, the Target Reward Assessment component evaluates their performance and provides feedback, which are further utilized to adaptively optimize intervention policy.

Method	Content quality		Behavior alignment	Self-consistency	Social capability
Method	BERTScore F1 $\uparrow$	BertSim $\uparrow$	Accuracy $\uparrow$	Accuracy $\uparrow$	Engagement $\uparrow$	Robustness $\uparrow$	Suitability $\uparrow$
Random	28.55 ±4.18	74.42 ±12.08	36.11 ±25.30	21.20 ±1.65	2.65 ±0.55	2.31 ±0.54	2.11 ±0.14
GLM4-light ( GLM et al., 2024 )	46.01 ±10.35	81.95 ±13.64	48.33 ±27.64	47.20 ±27.64	2.91 ±0.40	2.70 ±0.57	57.64 ±0.50
Llama-3-8B-Instruct ( AI@Meta, 2024 )	46.32 ±13.07	85.22 ±6.81	52.36 ±26.21	45.60 ±25.95	3.16 ±0.49	3.51 ±0.50	71.85 ±0.45
Qwen2.5-0.5B-Instruct ( Team, 2024 )	46.64 ±23.99	81.38 ±9.03	47.78 ±22.74	24.30 ±19.52	2.95 ±0.45	2.43 ±0.57	47.22 ±0.50
Qwen2.5-3B-Instruct	48.26 ±12.57	85.91 ±6.35	60.56 ±26.87	40.40 ±26.98	3.17 ±0.47	2.65 ±0.61	52.41 ±0.50

Table 2. Micro-level performance of social simulation across different backbones and PolicySim on the TwiBot-20 dataset. Underline indicates the best result and results are averaged over five runs with standard deviations.

Method	Objective 1			Objective 2
Method	Stance	Toxicity $\downarrow$	Cross interactions $\uparrow$	Misinformation ratio $\downarrow$
Origin	0.014 (0.47)	0.0556	0.04	40%
[1pt/1pt] $\epsilon$ -greedy	0.184 (0.42)	0.0426	0.14	26%
UCB	0.026 (0.34)	0.0628	0.50	30%
[1pt/1pt] PolicySim	0.376 (0.48)	0.0386	0.56	24%

Table 3. Performance of different intervention policies across different objectives. Underline indicates the best performance.

Figure 3. Mean and Std. of stance score by different intervention policies under trigger news. w/ IP. and w/o IP. indicate whether an intervention policy is applied.

Figure 4. Scalability of PolicySim on the TwiBot-20 dataset, showing linear runtime growth with agent scale ( $r=0.9904$ ).

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

	# Node	# Edges	# of tweets	Max Degree	Average Degree	Density	Average Cluster Modularity	Average Length of tweet
Value	924	302	20,061	14.0	0.65	3.54e-4	0.96	115.61

Appendix

Table 4. Dataset statistics for Twibot-20 sampled dataset.

Temperature ( $\tau$ )	0.4	0.5	0.6	0.7	0.8	0.9	1.0
PolicySim	0.7584	0.7649	0.7614	0.7632	0.7675	0.7769	0.7523

Appendix

Table 5. Robustness evaluation of PolicySim under varying temperature parameters (

\tau

) on Hunyuan-Lite. Moderate temperatures (

\tau=0.9

) achieve the best balance between generation diversity and semantic coherence.

	Engagement $\uparrow$	Robustness $\uparrow$	Suitability $\uparrow$
Qwen2.5-3B-Instruct	3.15±0.48	2.61±0.56	52.08±0.50
[1pt/1pt] PolicySim	3.28±0.53	2.68±0.53	69.86±0.46

Appendix

Table 6. Micro-level simulation result of PolicySim on Weibo dataset.

	Avg. Reward	Avg. Toxicity	LLM agents stances	ABM stances
PolicySim	0.3661	0.0392	0.5064( 0.3141 )	-
HiSim Hybrid w/ RA	0.1596	0.0487	0.5340(0.1226)	0.5067(0.1726)
HiSim Hybrid w/ Lorenz	0.1724	0.0471	0.5296(0.1187)	0.5016(0.1789)

Appendix

Table 7. Quantitative comparison between PolicySim and HiSim-based baselines under identical settings. PolicySim achieves the highest reward and lowest toxicity, demonstrating superior intervention optimization performance.

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 242 papers

LLM-Agent-based Social Simulation for Attitude Diffusion
Apr 4, 2026Deepak John RejiLarge Language Model-Based Social Simulation Computational Social Science
Department of Sociology, University of Limerick, Limerick, Ireland
From Plausible to Causal: Counterfactual Semantics for Policy Evaluation in Simulated Online Communities
Apr 5, 2026Agam Goyal, Yian Wang, Eshwar Chandrasekharan +1Counterfactual Evaluation Large Language Model-Based Social Simulation
University of Illinois Urbana-Champaign, Urbana, Illinois, USA
Stop Drawing Scientific Claims from LLM Social Simulations Without Robustness Audits
May 17, 2026Jinyi Ye, Lei Cao, Ding Chen +1Human Behavior Simulation Agent-Based Modeling
Thomas Lord Department of Computer Science, University of Southern California · Annenberg School for Communication and Journalism, University of Southern California · Marshall School of Business, University of Southern California

Apr 4, 2026J/K move · Enter open · S save

LLM-Agent-based Social Simulation for Attitude Diffusion

Deepak John Reji

Department of Sociology, University of Limerick, Limerick, Ireland

Large Language Model-Based Social Simulation Computational Social Science Human Behavior Simulation Agent-Based Modeling

Open paper AskPDF arXiv

This paper introduces discourse_simulator, an open-source framework that combines LLMs with agent-based modelling. It offers a new way to simulate how public attitudes toward immigration change over time in response to salient events like protests, controversies, or policy debates. Large language models (LLMs) are used to generate social media posts, interpret opinions, and model how ideas spread through social networks. Unlike traditional agent-based models that rely on fixed, rule-based opinion updates and cannot generate natural language or consider current events, this approach integrates multidimensional sociological belief structures and real-world event timelines. This framework is wrapped into an open-source Python package that integrates generative agents into a small-world network topology and a live news retrieval system. discourse_sim is purpose-built as a social science research instrument specifically for studying attitude dynamics, polarisation, and belief evolution following real-world critical events. Unlike other LLM Agent Swarm frameworks, which treat the simulations as a prediction black box, discourse_sim treats it as a theory-testing instrument, which is fundamentally a different epistemological stance for studying social science problems. The paper further demonstrates the framework by modelling the Dublin anti-immigration march on April 26, 2025, with N=100 agents over a 15-day simulation. Package link: https://pypi.org/project/discourse-sim/

Abstract

Large Language Model-Based Social Simulation

Explore similar work

LLM-Agent-based Social Simulation for Attitude Diffusion

From Plausible to Causal: Counterfactual Semantics for Policy Evaluation in Simulated Online Communities

Stop Drawing Scientific Claims from LLM Social Simulations Without Robustness Audits