Discourse Simulator

LLM-augmented Agent-Based Social Simulation for modelling public discourse dynamics following a critical real-world event.

Agents observe live news, reason about it, post to a simulated social network, and update multidimensional beliefs through peer influence and news salience, all powered by a local Ollama LLM.

---

Installation

1. Install discourse-sim

pip install discourse-sim

2. Install Ollama

discourse_sim runs LLMs locally via Ollama. You need to install it separately.

macOS

brew install ollama

Or download the desktop app from ollama.com/download.

Linux

curl -fsSL https://ollama.com/install.sh | sh

Windows

Download and run the installer from ollama.com/download.

3. Start the Ollama server

ollama serve

On macOS with the desktop app, this starts automatically. On Linux/Windows you may need to run this in a separate terminal or set it up as a service.

4. Pull the default model

ollama pull mistral:7b-instruct-q4_0

This is the default model (~4 GB). You can use any other Ollama model by passing ollama_model="..." to DiscourseSimulation.

Lighter alternatives (if disk space is limited):

ollama pull mistral          # ~4 GB  - good balance
ollama pull llama3.2         # ~2 GB  - faster, slightly less capable
ollama pull phi3             # ~2 GB  - Microsoft, very efficient

5. Verify everything works

ollama run mistral:7b-instruct-q4_0 "Hello"

You should see a response. If you do, you are ready to run simulations.

All dependencies: (make sure if everything is installed correctly)

pip install langchain-core langchain-community langchain-ollama \
            duckduckgo-search networkx numpy tqdm pandas openpyxl

---

Quick Start

from discourse_sim import DiscourseSimulation

sim = DiscourseSimulation(
    critical_event=(
        "In late April 2025, Dublin saw a large anti-immigration march "
        "from the Garden of Remembrance to the Custom House, with thousands "
        "protesting immigration levels and housing pressure..."
    ),
    event_date="2025-04-26",
    topic="immigration in Ireland",
    n_agents=100,
    n_days=15,
)

sim.run()
df = sim.to_dataframe()   # one row per agent per day
df.to_excel("output.xlsx", index=False)

---

All Parameters

Required

Parameter	Type	Description
critical_event	str	Plain-text description of the event injected into every agent prompt on every day
event_date	str	ISO date "YYYY-MM-DD" - this becomes Day 0
topic	str	Subject domain, e.g. "immigration in Ireland". Anchors search queries and scoring prompts

Agent Population

Parameter	Type	Default	Description
n_agents	int	100	Number of synthetic agents
n_days	int	15	Simulation duration in days
agent_distribution	dict	Ireland 2025 empirical split	Proportions of each agent kind - must sum to 1.0
kind_priors	dict	DEFAULT_KIND_PRIORS	Prior distributions for belief/attitude initialisation per kind

Default distribution (Ireland 2025):

{"centrist": 0.45, "far_right": 0.20, "pro_imm": 0.25, "media": 0.10}

Timeline

Parameter	Type	Default	Description
timeline	dict[int, str]	None	Optional explicit daily news entries keyed by 0-based day index. Missing days auto-generated. If None, all days use live search.

Example:

timeline={
    0: "[VERIFIED] The march took place at the Garden of Remembrance...",
    4: "[VERIFIED] Government announced a deportation flight...",
}

Any day not in the dict falls back to an auto-generated entry. Agents search live news on every day regardless of timeline.

LLM & Network

Parameter	Type	Default	Description
ollama_model	str	"mistral:7b-instruct-q4_0"	Any locally pulled Ollama model
temperature	float	0.75	Post generation temperature
use_llm_scoring	bool	True	If False, uses keyword heuristic for attitude scoring
network_k	int	6	Watts-Strogatz nearest-neighbour count
network_p	float	0.3	Watts-Strogatz rewiring probability
network_seed	int	42	Reproducibility seed for network and agents

Belief Update Keywords

Parameter	Type	Default	Description
threat_keywords	list[str]	See config.py	Words in daily news that trigger security_threat_belief increase and exposure accumulation
humanitarian_keywords	list[str]	See config.py	Words that trigger humanitarian_belief increase

Override these when simulating non-immigration topics:

threat_keywords=["eviction", "homeless", "unaffordable", "crisis"],
humanitarian_keywords=["social housing", "rights", "family", "support"],

---

Outputs

sim.run()

# Single DataFrame - one row per agent per day (recommended for analysis)
df = sim.to_dataframe()

# All three DataFrames
dfs = sim.to_dataframes()
dfs["history"]   # one row per timestep - aggregate metrics
dfs["agents"]    # one row per agent per timestep - full belief state
dfs["messages"]  # one row per agent per timestep - posts + beliefs merged

Column reference - df_messages

Column	Description
t / date	Day index and calendar date
agent_id / kind / quirk	Agent identity
message	The generated social media post
interpreted_score	LLM-scored stance [-1, +1]
attitude	Agent attitude at end of day t
mood	Agent mood at end of day t
exposure	Cumulative threat narrative exposure
economic_threat_belief	Economic threat sub-belief
cultural_threat_belief	Cultural threat sub-belief
security_threat_belief	Security threat sub-belief
humanitarian_belief	Humanitarian weight sub-belief
avg_attitude	Population mean attitude that day
polarization	Mean edge-level attitude divergence
news_summary	First 120 chars of daily news entry

---

Custom Agent Kinds

Add any agent kind by extending DEFAULT_KIND_PRIORS:

from discourse_sim import DiscourseSimulation
from discourse_sim.config import DEFAULT_KIND_PRIORS

custom_priors = {
    **DEFAULT_KIND_PRIORS,
    "nationalist": {
        "attitude":        (0.55, 0.95),
        "economic_threat": (0.2,  0.6),
        "cultural_threat": (0.7,  1.0),
        "humanitarian":    (-0.6, 0.0),
        "openness":        (0.1,  0.35),
        "emotional_react": (0.5,  0.85),
    },
}

sim = DiscourseSimulation(
    ...,
    agent_distribution={"centrist": 0.35, "nationalist": 0.30, "pro_imm": 0.35},
    kind_priors=custom_priors,
)

---

Architecture

discourse_sim/
├── __init__.py          # Public API: DiscourseSimulation
├── core.py              # DiscourseSimulation class - orchestrator
├── config.py            # SimConfig dataclass - all parameters + validation
├── agents/
│   └── agent.py         # Agent dataclass + make_agents() factory
├── tools/
│   └── search_tools.py  # make_tools() - search, memory, sentiment
├── timeline/
│   └── timeline.py      # Timeline class - user-supplied + auto-generated
├── simulation/
│   └── engine.py        # generate_message(), score_message(), update_beliefs(), run_loop()
└── utils/
    └── dataframes.py    # build_dataframes() - history/agents/messages

Per-day flow for each agent:

OBSERVE  →  search_event_news(query)          # live DuckDuckGo
         →  recall_agent_memory(posts_json)   # last 5 posts

THINK    →  build prompt: event + profile + memory + search result + today's news

ACT      →  LLM call → post text (max 40 words)

SCORE    →  LLM call (temp=0.0) → float in [-1, +1]

UPDATE   →  news salience → sub-beliefs
         →  peer pull     → attitude pressure
         →  mood shock    → affective state
         →  inertia       → resistance to change
         →  composite     → attitude(t)

---

Examples

File	Description
examples/dublin_march.py	Full reproduction of the original Dublin April 2025 experiment
examples/minimal_usage.py	Minimal usage - no timeline, live search only
examples/custom_agent_kinds.py	Adding a custom nationalist kind with its own priors

---

Citation

If you use this package in research, please cite:

@software{discourse_sim,
  title  = {discourse\_sim: LLM-Augmented Agent-Based Social Simulation of Discourse Dynamics},
  author = {dreji18},
  year   = {2026},
  note   = {Python package}
}