llm-society 0.3.2

LLM-driven agent-based information diffusion simulation

LLM Society — LLM-driven Information Diffusion

A Python package to simulate information diffusion with LLM-based agent conversations. It supports metric scoring in [0,1], segments-based personas, interventions, polished visualizations, and a simple CLI.

Links

• Tutorial (LLM network, segments, interventions, custom graphs, export): docs/TUTORIAL.ipynb

Features

• Segment-based persona configuration (proportions, flexible trait specs; optional segment names)
• Random network generation with tie strengths, or use your own NetworkX graph
• LLM-driven conversations and numeric scoring in [0,1] (metric-based), or simple/complex contagion modes
• Tie strengths influence edge sampling, talk probability, conversation depth, and can grow/decay over time (even for non-adjacent pairs via all-pairs mode)
• Optional agent memory to keep recent utterances in-context for longer-term continuity
• Multi-metric scoring per topic (e.g., credibility, emotion, action intent) with user-defined prompts and joint JSON outputs
• Interactive dashboards (Plotly/Bokeh) for rapid, shareable analysis
• Group plots (by traits or by segment), intervention effect plots, centrality plots, animations
• YAML/JSON config + CLI; exporting history/scores/conversations

Installation

1. Python 3.10+
2. Install

pip install -r requirements.txt

3. Provide OpenAI key (LLM mode)

export OPENAI_API_KEY=<YOUR_OPENAI_API_KEY>
# or use a file (first line)
echo "<YOUR_OPENAI_API_KEY>" > api-key.txt

Quickstart (Notebook)

from llm_society.api import network
from llm_society.viz import set_theme

set_theme()
net = network(
  information="5G towers cause illness.",
  n=20, degree=4, rounds=10,
  talk_prob=0.25, mode="llm", complex_k=2, rng=0
)
net.simulate()             # conversations, score updates, summaries
net.plot(type="final")
net.plot(type="centrality", metric="degree", show_exposure=False)

Plotting

• final: final node scores heatmap on the graph
• coverage: coverage (exposed & score>0) over time
• group: mean score by group (by="traits" with attr in segments' traits; or by="segment")
• centrality: centrality vs final score; optionally add exposure panel via show_exposure=True
• intervention: mean score over time with intervention marker; optionally group by traits
• animation: animated score evolution

Advanced Capabilities

• Grouping
  • Traits: net.plot(type="group", by="traits", attr="political")
  • Segment: net.plot(type="group", by="segment", groups=["High-Dem", "High-Rep"])
• Interventions

  net = network(..., intervention_round=6, intervention_nodes=[0,1,2], intervention_content="Be skeptical...")
  net.simulate()
  net.plot(type="intervention", attr="political", groups=["Democrat","Republican"])
  

• Custom Graph Personas
  • If you pass graph=G and omit segments, personas are built from node attributes (gender, race, age, religion, political; others go to extra).
• All-pairs conversations (LLM mode)
  • Set conversation_scope="all_pairs" (or CLI --conversation-scope all_pairs) to allow any node pair to chat.
  • Pairs without edges start at weight 0 but still get a small selection chance; repeated conversations strengthen their tie and add the edge into the network.
• Multi-metric scoring
  • Define metrics (list of {id,label,prompt}) so each conversation returns structured JSON with coordinated scores for both speakers.
  • The first metric acts as the "primary" score used by legacy APIs/plots; additional metrics are stored in history[*].scores_multi and can be visualized via net.plot(..., metric="emotion").
• Intervention-only runs
  • Leave information="" and configure intervention_round, intervention_nodes, and intervention_content.
  • Agents chat casually until the intervention round starts, after which conversations probabilistically focus on the treatment content.
• Agent memory
  • Set memory_turns_per_agent > 0 (e.g., 4–8) to inject that many recent utterances (self + partners) into each agent’s system prompt so they can recall past exchanges.

Exporting

net.export(
  history_csv="history.csv",
  scores_csv="scores_by_round.csv",
  conversations_jsonl="conversations.jsonl",
)

# interactive dashboard inside notebooks
fig = net.dashboard(engine="plotly", attr="political", metric="credibility")
fig

# or save to HTML manually
html = net.dashboard(engine="plotly", attr="political", metric="credibility", to_html=True)
Path("dashboard.html").write_text(html, encoding="utf-8")

CLI

# write an example config
llm-society --write-example-config my-config.yaml
# run with a config
llm-society --config my-config.yaml
# or run fully via flags
llm-society \
  --information "Claim text" --n 20 --degree 4 --rounds 10 \
  --depth 0.6 --depth-max 6 --edge-frac 0.5 --conversation-scope all \
  --seeds 0,1 --talk-prob 0.25 --mode llm --complex-k 2 --rng 0

Configuration (overview)

• Core: n, degree, rounds, depth (0–1), max_convo_turns, edge_sample_frac
• Seeds: seed_nodes, seed_score
• Info/LLM: information (may be blank if an intervention is configured), talk_information_prob, model, metric_name, metric_prompt
• Metrics: optional metrics=[{id,label,prompt}, ...] to request multi-dimensional scoring (first metric remains the default for legacy APIs)
• Modes: contagion_mode in {llm, simple, complex}, complex_threshold_k
• Conversation scope: conversation_scope in {edges, all}, pair_weight_epsilon (minimum sampling weight boost for zero-tie pairs)
• Memory: memory_turns_per_agent (0 disables memory)
• Personas: persona_segments (with proportion, traits, optional name)

License

MIT