Agent-Based Modelling Meets Generative AI in Social Network Simulations

1 Introduction

• Manually configuring diverse interactions and information flows can oversimplify agents and social environments, introducing researcher bias and reducing fidelity to complex human behavior.The introduction identifies collective decision-making as a particular challenge.Source: §1 · p. 2
• The paper asks whether LLM-agents capture users’ interests and form communities or echo chambers, while also investigating their susceptibility to recommendation strategies.The introduction identifies personality traits and interests as open questions alongside community-level behavior.Source: §1 · pp. 2–3
• The proposed framework profiles users from their posts, then uses reasoning and interaction modules to model actions and content exposure in a social network.The interaction module supports preference-based, popularity-based, and random recommendations.Source: §1 · pp. 2–3
• LLM-agents mirror real users’ linguistic patterns and political leaning, form homogeneous ideological groups, and respond to recommendation strategies that promote engagement and echo chambers.The study uses Twitter data from the 2020 US election.Source: §1 · p. 3

2.1 Agent-Based Modelling for Social Media Simulation

• Traditional social-media ABM can require detailed calibration and predefined rules, leaving outcomes sensitive to assumptions and potentially biased representations of complex dynamics.The paper proposes generative agents to reduce reliance on explicit parameterization and predefined rules.Source: §2.1 · p. 4

2.2 Generative Agents

• Prior generative-agent studies report credible individual and social behavior, while this work extends social-media simulation beyond individual properties to network features.The paper contrasts its focus on personality and political interests, customizable exposure, and network formation with S3 [9].Source: §2.2 · p. 4
• The framework’s architecture pairs characterization from real users’ posts with simulation, where agents make contextual decisions and access recommended content.Decisions are represented as a Choice-Reason-Content triple.Source: Figure 1 · p. 5

3 Methodology

• The framework has complementary characterization and simulation phases for initializing agents and executing their decisions and interactions.The methodology describes the phases as the framework’s two main components.Source: §3 · p. 5

3.1 Characterization Phase

• Characterization uses users’ original posts to infer personality and interests, avoiding a priori parameters for agent profiles.The authors argue these traits represent engagement style and topics of genuine interest.Source: §3.1 · p. 5

3.2 Simulation Phase

• Agents can generate original content, reshare others’ content, or remain inactive; the LLM chooses actions and content based on traits, interests, and other agents’ behavior.Each response is structured as a Choice-Reason-Content triplet.Source: §3.2 · p. 6
• The Interaction Module uses RAG to retrieve relevant stored actions and content, enriching prompts with context for agents’ decisions.The vector database is updated with agent actions and published content, and retrieval operates as a recommendation system.Source: §3.2 · pp. 6–7, Figure 2 · p. 7
• The simulation queries agents sequentially in round-robin cycles, while its recommendation strategies are simplified approximations rather than actual methods such as collaborative filtering.Each iteration ends after all agents have been queried, then the next begins.Source: §3.2 · pp. 7–8

4.1 Experimental setup

• The study uses Dolphin 2.1 Mistral 7B and runs each simulation for 10 iterations, taking approximately 15 hours per run.Experiments used an Intel Core i7-11800H, 16 GB RAM, and an NVIDIA GeForce RTX 3060 Laptop GPU.Source: §4.1 · p. 8

4.2 Dataset

• The dataset contains over 12 million election-related tweets from 1.1 million users, collected from June to December 2020; 100 annotated users instantiate the agents.The 100 users include 73 Republicans and 27 Democrats; accounts flagged as bots were excluded, and original tweets were used to infer personality and interests.Source: §4.2 · p. 8

4.3 Characterizing LLM-agents vs. Real Users’ Interests (RQ1)

• LLM-agents reproduce the main topics and right-skewed keyword-frequency pattern of real election discussions, supported by a Mann–Whitney test (p-value< 0.05).The comparison uses extracted keywords from original content; the figure plots keyword occurrence counts against the probability of each frequency.Source: §4.3 · p. 9, Figure 3 · p. 10, Table 1 · p. 10
• At least 86% of agents retain their impersonated users’ political orientations across recommendation strategies, while shifts typically involve nonpartisan users.The authors report that exposure to diverse ideas may blur the political orientations of these shifting agents.Source: §4.3 · pp. 9–10, Table 2 · p. 10
• Under preference-based recommendations, median self-similarity is 0.274 for real users and 0.531 for LLM-agents; intra-cluster similarity shows minimal differences among the three distributions.Random recommendations better approximate real-user self-similarity, while preference-based recommendations better align with real-user intra-cluster similarity.Source: §4.3 · p. 11, Figure 4 · p. 12, Figure 5 · p. 12

4.4 Community Formation and Echo Chambers Among LLM-Agents (RQ2)

• Resharing falls to ≤8.4% with random recommendations versus 50.6% with preference-based recommendations, while non-interactions are higher with preference-based recommendations (10.5% versus ≤2.5%).The authors attribute the non-interaction difference to occasional irrelevant recommendations; agents using random recommendations publish original content about 90% of the time.Source: Table 2 · p. 10, §4.4 · p. 12
• Preference-based recommendations produce stronger homophily and more controversy, while random recommendations are not statistically homophilic and mitigate echo-chamber formation.Preference-based networks have 10% fewer inter-cluster edges than the non-homophily threshold; random-recommendation networks have 4% more inter-cluster edges than preference-based networks.Source: §4.4 · pp. 13–14, Table 3 · p. 14
• Agents’ directed resharing interactions form political clusters, and preference-based recommendations reduce connections between those clusters.The interaction graph represents agents as nodes and resharing from a publisher to a resharing agent as a directed edge.Source: §4.4 · pp. 12–13, Table 3 · p. 14

• LLM-agents maintained users’ political orientations and community themes, and tended to cluster with like-minded peers, especially under preference-based recommendations.These findings concern Twitter discussions surrounding the 2020 US election.Source: §5 · p. 14
• The framework combines a Characterization Module for inferring user traits with an Interaction Module that applies recommendation strategies through retrieval-augmented generation (RAG).Source: §5 · p. 14
• Future work will explore more sophisticated recommendation algorithms, broader agent actions such as following and liking posts, and multiple LLMs to analyze potential model biases.Source: §5 · p. 14
• The authors also identify potential applications beyond social media simulation, including epidemiology and business planning.Source: §5 · pp. 14–15