Research map / Multi-agent systems and games / LLM agents
Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces GovSim, a multi-agent common-pool resource simulation for LLMs, finding only the strongest models sustain resources and that universalization prompting and communication improve sustainability.
Problem
Whether LLM agents can cooperate to sustainably manage shared resources in multi-turn, multi-agent social dilemmas, where greedy actions lead to collapse.
Approach
A generative-agent simulation platform with three common-pool resource scenarios (fishery, pasture, pollution) where agents harvest, then discuss in natural language; agents are prompted with a universalization statement ("what if everyone does that") to improve long-term sustainability.
Key results
- Highest survival rate across models is 53.3% for GPT-4o in the default experiment, and most open-weights models achieve 0% survival.
- Universalization prompting increases average survival time by 4 months (t-test p<0.001) and efficiency by 24%.
- Disabling communication increases resource over-usage by 22% (p<0.001).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add universalization reasoning to agent memory | LLM agents in GovSim · GovSim (excluding Claude-3 Opus; combinations not already at maximum survival time) | survival time: +4 months | Helped |
| Add universalization reasoning to agent memory | LLM agents in GovSim · GovSim (excluding Claude-3 Opus; combinations not already at maximum survival time) | total gain: +29 units of shared resource | Helped |
| Add universalization reasoning to agent memory | LLM agents in GovSim · GovSim (excluding Claude-3 Opus; combinations not already at maximum survival time) | efficiency: +24% | Helped |
| Remove communication between agents (ablation) | GPT-4o, GPT-4-turbo, Claude-3 Opus, and Qwen-110B agents · GovSim | over-usage: +22% | Hurt |
| Introduce a greedy newcomer unfamiliar with the community norm | GPT-4o agents · GovSim (three scenarios) | survival rate: 53.3 -> 33.3 | Hurt |
| Introduce a greedy newcomer unfamiliar with the community norm | GPT-4o agents · GovSim (three scenarios) | survival time: 9.3 -> 6.6 | Hurt |
| Introduce a greedy newcomer unfamiliar with the community norm | GPT-4o agents · GovSim (three scenarios) | total gain: 66.0 -> 34.8 | Hurt |
| Introduce a greedy newcomer unfamiliar with the community norm | GPT-4o agents · GovSim (three scenarios) | efficiency: 55.0 -> 31.3 | Hurt |
| Introduce a greedy newcomer unfamiliar with the community norm | GPT-4o agents · GovSim (three scenarios) | equality: 94.4 -> 71.7 | Hurt |
| Introduce a greedy newcomer unfamiliar with the community norm | GPT-4o agents · GovSim (three scenarios) | over-usage: 10.8 -> 15.7 | Hurt |
Surprising findings
- Greedy newcomer perturbation drops GPT-4o survival rate from 53.3% to 33.3%, and inequality rises.
- Smaller models such as Llama-3-8B never sustain any resource past the first month.
Limitations
- Scenarios are simplified compared to real-world resource management.
- Results depend on greedy decoding and nondeterministic API inference.
- Only tests with a fixed population of five agents and 12 months.
Experimental setup
Datasets: procedurally generated sub-skill reasoning tests (150 problems per test)
Benchmarks: GovSim
Metrics: survival rate survival time total gain efficiency equality (Gini) over-usage
Baselines: GPT-4o GPT-4-turbo Claude-3 Opus Qwen-110B Llama-3-70B GPT-4 Claude-3 Sonnet
Builds on: Generative Agents Tragedy of the Commons Machiavelli benchmark GTBench
Closest papers by meaning
- More Capable, Less Cooperative? When LLMs Fail at Zero-Cost Collaboration (ICML 2026 · similarity 0.90)
- Many LLMs Are More Utilitarian Than One (NeurIPS 2025 · similarity 0.89)
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
- Shapley-Coop: Credit Assignment for Emergent Cooperation in Self-Interested LLM Agents (NeurIPS 2025 · similarity 0.89)
- Moral Alignment for LLM Agents (ICLR 2025 · similarity 0.88)
- Evolving Interpretable Constitutions for Multi-Agent Coordination (ICML 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.