Recognition: no theorem link
LLM-Agent-based Social Simulation for Attitude Diffusion
Pith reviewed 2026-05-13 16:43 UTC · model grok-4.3
The pith
An open-source package lets large language models simulate how attitudes toward immigration spread through social networks after real events.
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
Core claim
The discourse_sim framework combines large language models with agent-based modeling so that generative agents placed in a small-world network topology can produce and interpret social media content while retrieving live news. This setup simulates how public attitudes toward immigration shift in response to salient events, treating the entire simulation as an instrument for testing sociological theories of belief evolution and polarization.
What carries the argument
The discourse_simulator package, which embeds LLM-powered generative agents in a small-world network linked to a live news system so agents can generate posts and revise opinions from real-world timelines.
If this is right
- Researchers can run controlled tests of how specific protests or policy debates alter belief distributions across the network.
- The framework supports incorporation of current event timelines instead of relying only on static rules.
- Simulations become repeatable instruments for exploring polarization mechanisms under different network conditions.
- The same structure can be reused to study attitude dynamics on other topics by changing the news feed and initial agent beliefs.
Where Pith is reading between the lines
- The approach could be tested by replaying older events with known survey outcomes to check whether simulated shifts match historical records.
- Adding agent heterogeneity in news consumption patterns might reveal subgroups that drive faster diffusion than the average network suggests.
- Extending the model to track belief consistency across multiple issues could show whether single-event shocks produce lasting cross-topic alignment.
Load-bearing premise
Large language models can generate and interpret social media content in ways that faithfully represent real human opinion dynamics and information spread within the modeled network.
What would settle it
Compare the direction and size of attitude change produced by the fifteen-day simulation of the Dublin march against independent post-event survey data or sentiment measurements from actual social media.
read the original abstract
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/
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper introduces discourse_simulator, an open-source Python framework that integrates LLM-based generative agents into a small-world network topology with live news retrieval to simulate multidimensional attitude diffusion and polarization in response to real-world events such as protests. It positions the tool as a theory-testing instrument for social science rather than a black-box predictor, and demonstrates it via a single 15-day run with N=100 agents modeling the April 2025 Dublin anti-immigration march.
Significance. If the framework's outputs can be validated against empirical data, it would offer social scientists a flexible instrument for testing theories of opinion dynamics that incorporates natural-language generation and current events, going beyond traditional rule-based agent-based models.
major comments (2)
- [Demonstration] Demonstration section: the N=100 agent, 15-day simulation of the Dublin march is described without any reported quantitative outputs (e.g., polarization metrics, sentiment trajectories, or volume of generated posts), without comparison to real social-media time series or post-event surveys, and without ablation or sensitivity analyses. This leaves the central claim that the framework produces realistic attitude diffusion without supporting evidence inside the manuscript.
- [Framework description] Framework description (architecture and epistemological stance sections): the assertion that discourse_sim functions as a 'theory-testing instrument' is not illustrated by any concrete example of hypothesis formulation, model calibration against sociological theory, or falsification procedure; the single run remains purely descriptive.
minor comments (2)
- [Implementation] The package link is given but no installation instructions, dependency list, or reproducibility script for the Dublin demonstration appear in the text.
- [Model specification] Notation for agent belief structures and network parameters is introduced without a consolidated table or explicit default values.
Simulated Author's Rebuttal
We thank the referee for their constructive comments, which help clarify how to better position the framework. We address each major point below and have revised the manuscript accordingly to provide supporting evidence and concrete illustrations.
read point-by-point responses
-
Referee: [Demonstration] Demonstration section: the N=100 agent, 15-day simulation of the Dublin march is described without any reported quantitative outputs (e.g., polarization metrics, sentiment trajectories, or volume of generated posts), without comparison to real social-media time series or post-event surveys, and without ablation or sensitivity analyses. This leaves the central claim that the framework produces realistic attitude diffusion without supporting evidence inside the manuscript.
Authors: We acknowledge that the current demonstration is primarily descriptive. In the revised manuscript we will add quantitative outputs from the N=100 run, including multidimensional polarization metrics (standard deviation and variance across attitude dimensions), daily sentiment trajectories, and total volume of generated posts. We will also include a limited sensitivity analysis on network parameters (e.g., rewiring probability) and LLM sampling temperature. While a full empirical comparison to social-media time series or post-event surveys lies beyond the scope of an introductory framework paper, we will add a qualitative discussion of alignment with publicly reported trends from the Dublin march coverage. revision: yes
-
Referee: [Framework description] Framework description (architecture and epistemological stance sections): the assertion that discourse_sim functions as a 'theory-testing instrument' is not illustrated by any concrete example of hypothesis formulation, model calibration against sociological theory, or falsification procedure; the single run remains purely descriptive.
Authors: We agree that an explicit worked example is needed. The revised manuscript will include a new subsection that formulates a testable hypothesis drawn from social identity theory (increased out-group bias following protest events), describes how initial agent beliefs can be calibrated against existing survey data on immigration attitudes, and outlines a falsification check (failure to observe expected polarization under high-clustering networks). This addition demonstrates the intended theory-testing workflow while keeping the paper focused on framework introduction rather than a complete empirical study. revision: yes
Circularity Check
No circularity in framework presentation
full rationale
The paper introduces an open-source software framework combining LLMs with agent-based modeling for attitude diffusion simulation. No mathematical derivations, equations, fitted parameters, or self-citation chains are present that reduce any output or prediction to the paper's own inputs by construction. The contribution is the Python package and its use as a theory-testing instrument, with the Dublin demonstration serving as an illustrative run rather than a closed-form result. This is self-contained as a tool-building paper with no load-bearing circular steps.
Axiom & Free-Parameter Ledger
axioms (2)
- domain assumption Small-world network topology models realistic social connections for opinion spread
- domain assumption LLM-generated posts and interpretations can stand in for real human social media behavior
Reference graph
Works this paper leans on
-
[1]
LLM-Agent-based Social Simulation for Attitude Diffusion Deepak John Reji Department of Sociology, University of Limerick, Limerick, Ireland reji.deepak@ul.ie Keywords: agent-based modelling; social network dynamics; opinion polarisation; large language models; attitude formation; computational social science Abstract This paper introduces discourse_simul...
work page 2025
-
[2]
Introduction We live in a world that is evolving rapidly technologically, socially, and politically. We have witnessed the evolution of AI, particularly the rise of autonomous systems, and the evolution of human society, and, at the same time, the unfolding of difficult realities: wars, humanitarian crises, natural disasters, and struggles for survival. I...
work page 2023
-
[3]
Background and Related Work 2.1 Generative Social Science and the ABM Tradition Epstein argued that agent-based computational models enable a distinctively "generative" approach to social science to explain a social phenomenon, one must be able to grow it from the bottom up through decentralised local interactions of heterogeneous autonomous agents. His m...
work page 1999
-
[4]
(Gilbert and Terna 2000). The two key challenges which limited the adoption of ABMs in social sciences were the oversimplification of Human Behaviour (Lack of Realism) and the lack of Empirical Grounding (Validation and Calibration). Traditional models were struggling to replicate human reasoning, storytelling, learning, and cognitive biases. The Cognitiv...
work page 2000
-
[5]
large anti-immigration protest in Dublin
Framework Design The package is a five-layer sequential pipeline. The user interacts exclusively with the discourse-simulator library, which orchestrates all internal components via SimConfig, which instantiates agents, the social network, LLM tools, and the news timeline in parallel. The engine executes the Observe-Think-Act (OTA) loop across T days × N ...
work page 2025
-
[6]
Uniform(-0.4, +0.4) pro_imm 25% Younger cohort, NGO workers, immigrant community; under-35s significantly more pro-immigration per ESRI surveys Uniform(-1.0, -0.3) far_right 20% ~52–59% of Irish adults say immigration 'too high' (LSE/ESRI 2024), but hard far-right identity is far narrower than anti-immigration sympathy Uniform(+0.5, +1.0) media 10% Journa...
work page 2024
-
[7]
requires the LLM to parse tool-call schemas and this fails for small instruction-tuned models like Mistral 7B. Besides the manual OTA loop decouples tool execution (Python) from reasoning (LLM), making the framework model-agnostic and reproducible. Phase 1: Observe (Tool Calls) Python directly invokes the tools and collects their outputs as observations: ...
work page 2025
-
[8]
Security threat and humanitarian beliefs each contribute 20%, with humanitarian beliefs exerting a negative (pro-immigration) pressure. 3.6 Full Attitude Update The final attitude update integrates all five components: attitude(t) = clip(inertia × attitude(t-1) + (1 - inertia) × (0.4 × own_score + 0.3 × peer_pull + 0.3 × belief_composite), -1, +1 ) Within...
-
[9]
• Chuang, Y.-S., Goyal, A., Harlalka, N. et al. (2023). Simulating Opinion Dynamics with Networks of LLM-based Agents. NAACL-HLT
work page 2023
-
[10]
• Törnberg, P ., Valeeva, D., Uitermark, J. et al. (2023). Simulating Social Media Using Large Language Models to Evaluate Alternative News Feed Algorithms. arXiv:2310.05984. • Gao, C. et al. (2023). S3: Social-network Simulation System with Large Language Model-Empowered Agents. arXiv:2307.14984. • Mou, X., Wei, Z. & Huang, X. (2024). Unveiling the Truth...
-
[11]
Immigration Discourse • Entman, R.M. (1993). Framing: Toward clarification of a fractured paradigm. Journal of Communication, 43(4), 51–58. • Boomgaarden, H.G. & Vliegenthart, R. (2009). How news content influences anti-immigration attitudes. European Journal of Political Research, 48(4), 516–542. • Esses, V .M., Medianu, S. & Lawson, A.S. (2013). Uncerta...
work page 1993
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.