Literature Review¶

Key Papers¶

Fontana et al. (2024/2025) — "Nicer Than Humans"¶

• Scenario: 100-round IPD, LLM agents vs. random opponents that cooperate with probability α (0.0 to 1.0)
• Communication: None
• Agent introduction: System prompt with PD rules, payoff matrix, 10-round history window
• Motivation: "Maximize your score"
• Key finding: LLMs cooperate ~79% vs. humans ~37%. Llama3 was exploitative.
• Validation: Meta-prompting — asked LLMs to explain the game before playing

Akata et al. (2025) — "Playing Repeated Games" (Nature)¶

• Scenario: Only 10 rounds, 144 different 2×2 games. LLM-vs-LLM, LLM-vs-policy, LLM-vs-humans (195 people)
• Communication: None. Used neutral labels (J/F, not Cooperate/Defect) to avoid bias
• Agent introduction: Cover stories, abstract framing. Temperature 0.0
• Motivation: Score-maximizing. Added Social Chain-of-Thought (SCoT) — predict opponent's action before deciding
• Key finding: GPT-4 permanently defects after a single betrayal. "The other player sometimes makes mistakes" framing restored cooperation. SCoT improved coordination.
• Limitation: 10 rounds too short for strategies to evolve

Pal et al. (2026) — "Strategies of Cooperation/Defection in Five LLMs"¶

• Scenario: 50 API calls per scenario (hypothetical), then 10-round verification games
• Communication: None. Neutral labels (L/R)
• Agent introduction: 9 framings (neutral, competitive, cooperative, etc.)
• Key finding: Each model has a distinct "strategic personality" — Claude = Forgiver, GPT-4o = GRIM, Llama = Generous TFT. Claude wins 12 of 15 tournament disciplines.

Guo (2023) & Phelps (2023) — Persona Studies¶

• Tested personality prompts ("fair," "selfish," "altruistic," "competitive")
• Personas translate to behavior, but agents struggle with conditional reciprocity
• Cooperation only stable when both agents are cooperative

Piatti et al. (2024) — "Cooperate or Collapse" (NeurIPS)¶

• Scenario: 5 LLM agents share a renewable resource pool (commons dilemma)
• Communication: Yes — free-form natural language between agents
• Key finding: Communication reduced overuse by 21%. "Universalization" reasoning improved sustainability.
• Only the strongest models (GPT-4o) could sustain cooperation

Poje et al. (2024) — Private Deliberation & Deception¶

• Scenario: Repeated games with "private agent" (hidden chain-of-thought)
• Communication: Public messages, but private reasoning could differ
• Key finding: Private agents with deception capability achieved higher long-term payoffs. Deployed deception only when strategically optimal.

Summary: Communication Effects¶

Additional Key Papers¶

Weng et al. (2025) — "Will Systems of LLM Agents Cooperate" (AAMAS) - LLMs generate IPD strategies as natural language → Python. Evolutionary dynamics (Moran processes). Cooperative strategies generally outperform aggressive ones, but "Self-Refine" prompting enhances aggressive strategy effectiveness — a safety concern.

Huynh et al. (2025) — "Understanding LLM Agent Behaviours via Game Theory" - LSTM-based strategy recognition (94% accuracy at identifying ALLC/ALLD/TFT/WSLS from trajectories). Payoff scaling across languages and personalities. Cross-linguistic effects significant.

Vallinder & Hughes (2025) — "Cultural Evolution of Cooperation among LLM Agents" (Google DeepMind) - 10 generations of LLM agents in Donor Game. Claude societies evolve more cooperative; GPT-4o populations become increasingly untrusting.

Jia et al. (2025) — "LLM Strategic Reasoning" (NeurIPS) - Behavioral game-theoretic evaluation via Truncated Quantal Response Equilibrium. GPT-o3-mini has greatest reasoning depth. Each model has distinct reasoning style.

Hao et al. (2026) — "Game-Theoretic Lens on LLM-based MAS" - Comprehensive survey organizing LLM MAS research through game theory's four core elements (players, strategies, payoffs, information).

Zhu (2025) — "Game Theory Meets LLM: Cybersecurity" - Theoretical monograph proposing "LLM-Nash games" and "reasoning-level equilibrium" — behavior emerging from prompt-based reasoning rather than analytical computation.

Key Gaps We Address¶

1. No one tested communication as a controlled variable in standard IPD with LLMs
2. No one tested protocol safeguards (MCP-like) in game-theoretic settings
3. No one used geometric/unknown horizon (changes optimal strategy fundamentally)
4. No one mapped personas to classical strategies using canonical baselines
5. No one published in IS venues — all prior work is CS/AI/psychology
6. No one built a reproducible web-based benchmark with UI for experiment design
7. No one systematically tested LLMs against the full canonical strategy set (ALLC, ALLD, TFT, GRIM, GTFT, WSLS) — Fontana used only random opponents, Akata only 3 policy agents
8. No one modeled adversarial personas as an experimental variable in LLM PD with canonical baselines — Phelps/Guo tested basic personas but without systematic strategy mapping