Misinformation Spread Model (LLM vs Rule-Based)

A Mesa agent-based model simulating how misinformation spreads across a social network when agents reason using an LLM versus fixed probability rules. Models how different personality types (believers, skeptics, and spreaders) produce group dynamics that no individual agent was programmed to exhibit.

Overview

Agents are placed on an Erdos-Renyi graph and exposed to a misinformation claim. Each step, agents decide whether to believe and share the claim based on their personality type and how many times they have previously encountered it. LLM agents accumulate a decision history across steps, allowing social pressure to build up over time.

The key behavior is the S-curve effect: LLM-driven skeptics initially resist the claim for several steps, but repeated exposure gradually tips them toward belief. This pattern that emerges purely from prompt context and not hard-coded thresholds.

Installation

git clone <your-repo-url>
cd misinformation_spread
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

LLM requirement: This model uses Ollama for local LLM inference. Install Ollama and pull the model before running:

ollama pull llama3

Running the Model

Static comparison (2-panel chart):

python run.py

Interactive dashboard:

solara run app.py

Parameters

• n_believers (default: 4) - number of believer agents. Tend to trust and share information easily; belief strengthens with repeated exposure
• n_skeptics (default: 3) — number of skeptic agents. Initially resistant; may gradually convince after many exposures
• n_spreaders (default: 2) — number of spreader agents. Share everything to maximize engagement regardless of truth
• connectivity (default: 0.3) — probability of a connection between any two agents in the Erdos-Renyi graph. Higher values produce denser networks with faster spread
• misinformation_claim (default: cold water digestion claim) — the claim being spread. Can be replaced with any string
• rng (default: 42) — random seed for reproducibility

Model Details

Agents:

• BelieverAgent — LLM agent that trusts information easily. Passes times_exposed and full decision history into the prompt each step, allowing accumulated exposure to strengthen belief over time.
• SkepticAgent — LLM agent that questions claims. Starts resistant but the accumulation mechanism means repeated exposures gradually reduce skepticism, producing delayed belief.
• SpreaderAgent — LLM agent that shares everything to maximize engagement. Almost always says yes, reinforced by seeing the claim spread further each step.
• RuleBasedBeliever, RuleBasedSkeptic, RuleBasedSpreader — identical network structure but fixed probability thresholds (80%, 15%, 95%). Used as a baseline to isolate the effect of LLM reasoning.

Architecture note:

ReActReasoning.plan() wraps all output in reasoning: / action: fields regardless of prompt instructions, making structured two-line responses unparseable. This model bypasses ReActReasoning entirely and calls litellm.completion() directly, which produces clean BELIEVE: yes / SHARE: yes output that can be reliably parsed.

Network:

Built using networkx.erdos_renyi_graph and wrapped in Mesa's Network discrete space. Misinformation is seeded by broadcasting from all spreaders to all agents at initialization. Each step, any agent with shared=True propagates the claim to all other agents.

Social pressure mechanism:

Each agent tracks times_exposed and a history list of past decisions. Both are passed into the LLM prompt each step so the model can reason about change over time rather than making independent per-step decisions.

Output

run.py produces 2 panels:

• Spread count over time for LLM vs rule-based agents
• Final convinced count by agent type (believers, skeptics, spreaders)

app.py produces an interactive dashboard with:

• Spread Over Time: live line chart comparing LLM vs rule-based S-curves
• Final Counts: bar chart of convinced agents by type at end of run
• Raw Data: step-by-step dataframes for both models

Project Structure

misinformation_spread/
├── agents.py           # LLM agent classes (Believer, Skeptic, Spreader)
├── model.py            # MisinformationModel and RuleBasedMisinformationModel
├── rulebased_agents.py # Rule-based agent classes for baseline comparison
├── run.py              # Run both models and generate static comparison chart
├── app.py              # Interactive Solara dashboard
├── readme.md           # This file
├── requirements.txt
├── tests/
│   └── test_model.py   # pytest suite (rule-based only, no LLM calls)
└── images/             # Output charts saved here

References

• Vosoughi, S. et al. (2018). The spread of true and false news online. Science.