add: LLM Prisoner's Dilemma — when agents reason about trust, not just strategy

[abhinavk0220]

What this is

An iterated Prisoner's Dilemma simulation where agents use LLM
Chain-of-Thought reasoning to decide whether to cooperate or defect
instead of following fixed strategies like tit-for-tat or always-defect.

Why this is scientifically interesting

The Prisoner's Dilemma has been studied for 70+ years because it
captures a fundamental tension in social systems: individual rationality
(defect for personal gain) conflicts with collective optimality
(cooperate for mutual benefit).

Classical ABM approaches this with fixed rules:

• Always defect : pure self-interest, Nash equilibrium
• Tit-for-tat : copy partner's last move
• Pavlov : win-stay, lose-shift

These strategies are elegant but behaviorally brittle. Real humans
don't follow fixed rules : they reason about context, history,
reputation, and intent.

LLM agents do the same:

"My partner cooperated last 3 rounds but defected once when I
defected first. They seem reactive rather than purely selfish.
Cooperating now signals good faith and likely leads to stable
mutual cooperation long-term."

This produces emergent social dynamics : trust building, strategic
betrayal, forgiveness, reputation management : that no fixed strategy
can replicate.

Technical implementation

• PrisonerAgent extends LLMAgent with CoT reasoning
• Each round agents receive their full interaction history as
  internal state : score, last action, partner's recent moves
• Random pairing each round prevents fixed-partner exploitation
• Payoff matrix applied after both agents decide independently
  (simultaneous decision : no sequential advantage)
• Tracks cooperation_rate and avg_score as model-level metrics

Payoff matrix

	Partner Cooperates	Partner Defects
You Cooperate	3, 3	0, 5
You Defect	5, 0	1, 1

Visualization

• 🟢 Green agents : cooperated last round
• 🔴 Red agents : defected last round
• Agent size scales with cumulative score
• Cooperation rate plot : tracks social trust over time
• Average score plot : tracks collective welfare over time

Setup

cp .env.example .env   # fill in your API key
pip install -r requirements.txt
solara run app.py

Supports Gemini, OpenAI, Anthropic, and Ollama via .env.example.

Reference

Axelrod, R. (1984). The Evolution of Cooperation. Basic Books.