Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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, Yang Yang

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.

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

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.
Paper Structure (32 sections, 12 equations, 4 figures, 7 tables, 1 algorithm)

This paper contains 32 sections, 12 equations, 4 figures, 7 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Simulation Sandbox Framework
  3. User Agent Module
  4. User Profile.
  5. User Behavior.
  6. Relation.
  7. User Stance.
  8. Memory.
  9. Agent Training and Planning
  10. Intervention Policy Module
  11. Intervention Policy
  12. Intervention Objective.
  13. Adaptive Intervention Policy
  14. Adaptive Interventions via RL
  15. Experiments
  16. ...and 17 more sections

Figures (4)

  • 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.
  • 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.
  • 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 ($r = 0.9904$).