Opinion Polarization in LLM-Based Social Networks — Simulation Code

Reference implementation for "Opinion Polarization in LLM-Based Social Networks: Manipulation and Mitigation." The framework simulates a directed social network of LLM agents that exchange natural-language posts and update their opinions, and evaluates adversarial manipulation strategies and mitigation mechanisms.

Requirements

• Python 3.9+
• networkx, numpy, requests, tqdm, python-louvain (community)
• networkit — only required for the hrg graph type
• An OpenAI API key (or other companies related APIs) in the OPENAI_API_KEY environment variable

The current simulation calls the OpenAI Chat Completions API (https://api.openai.com/v1/chat/completions, model gpt-4.1-mini) for both post generation and opinion updates.

Files

File	Role
Runner.py	It's the Runner :), sweeps over manipulation/mitigation settings.
Run_LLM_using_API.py	Core simulation (Simulation class + main()): LLM calls and the opinion-dynamics loop.
Prompts.py	Post-generation and response-generation prompts (paper Appendix A).
AA.py	Manipulator (adversarial agent) selection: random / degree / betweenness / community-aware greedy.
CA.py	Reactive moderator selection (same four strategies, restricted to the moderator pool).
HRG_Graph.py	Synthetic Hyperbolic Random Graph generator (requires networkit).
SBM_Graph.py, SmallSBM_Graph.py	Stochastic block-model generators.
LLM_Utils.py	Graph builders (Reddit/Twitter/Facebook), opinion initialisation, metrics.

Terminology

• AA — adversarial agents (manipulators, the set A in the paper). --AA_type selection strategy; --AA_k budget k_A; --AA_level persistence (weak = susceptible, delta = 0.5; strong = persistent, delta = 1).
• CA — counter agents / mitigations (--CA_type). Reactive: moderator (neutral) and contrarian. Proactive: Broadening_social_ties, Distributed_activity_boost, Active_cross_checking_from_feeds, Active_cross_checking_from_zero, Resistance_extreme_content. --CA_param is the moderator selection strategy; --CA_k is the moderator budget k_M (or the proactive intensity parameter, depending on CA_type).

Model summary

Opinions live in [-1, 1] internally and are mapped to an integer percentage [0, 100] (0 = against, 100 = favor) when shown to the model, which the paper found to be the most stable representation. Each step: a node is activated with probability proportional to its activeness, it generates a post via the LLM, the post is broadcast to its outgoing neighbours, and each recipient updates its opinion via the LLM conditioned on its previous opinion and stubbornness.

Manipulators are initialised to an extreme opinion (+/-1) aligned with their initial leaning, and their activeness and stubbornness are increased by delta (min(1, x + delta)). Persistent manipulators become fully stubborn and remain fixed; susceptible manipulators can still update over time.

Reactive moderators (activeness 1) are drawn from a moderator pool (M_pool = V \ A, 20% of non-manipulator nodes). Neutral moderators hold a fixed opinion of 0; contrarian moderators post the opposite of their neighbourhood's dominant opinion, o_v = -sign(sum_{u in Gamma+} o_u).

Two objective measures are reported (paper Sec. 3.2): polarization = variance of opinions, extremization = mean absolute opinion.

Running

export OPENAI_API_KEY=sk-...
python Runner.py --graph_type twitter --topic "Support Remote Work" --iterations 2000

Runner.py runs three blocks per seed: (1) a no-manipulation baseline, (2) manipulation-only sweeps, and (3) mitigation (and manipulation+mitigation) sweeps. Results are pickled per run under results-* directories, each containing opinion_history, stubbornness, variance, final_opinions, post_history, and update_history.

Datasets

• Synthetic graphs (hrg, sbm, smallsbm, small) are generated in code.