GovSim: Emergence of Sustainable Cooperation in a Society of LLM Agents

Giorgio Piatti, Zhijing Jin, Max Kleiman-Weiner, Bernhard Schölkopf, Mrinmaya Sachan, Rada Mihalcea

Five LLM agents share a common-pool resource and must agree on limits to avoid collapse. A replication of Piatti et al. (NeurIPS 2024) across three scenarios.

paper source REPS package

game theory tragedy of the commons

Resource survival and outcomes

Fishing, baseline.

Original paperReplicationReplication run

Original paper: Table 4 (Piatti et al., 2024).

Replication: 50 runs from this databank, listed under Data; served ids shown as the paper’s.

About this paper

Five agents are fishermen sharing one lake. Each month every agent decides, privately, how many tons of fish to catch. The lake holds at most 100 tons and doubles whatever is left, so as long as the group takes no more than half, the lake lasts forever. Take more and it shrinks until nothing is left.

After each harvest the agents meet in a town hall and talk. A run lasts up to 12 months and ends early if the lake collapses, so months survived is the headline result. Harvest and equality say how well the group did while it lasted.

Baseline: the setup above.
Universalization: each agent is also told what would happen if everyone took the same amount.
No discussion: the town hall is removed.
Outsider: one agent is replaced by a newcomer who takes more than their share.

The same rules are retold as sheep farming on a shared pasture and pollution from factories on a river. Only the story changes.

Piatti et al., Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents, NeurIPS 2024.

Figure 1: a community of AI agents shares resources in pasture, fishery, and pollution scenarios, leading to cooperation or resource collapse. — **Figure 1.** GovSim’s three resource-sharing scenarios. From Piatti et al. (2024), CC BY-SA 4.0.

Figure 2: the fishing agent’s monthly cycle of harvesting, discussion, and reflection. Example prompts combine simulation rules and memories to choose a catch, speak in a group, and record insights for future decisions. — **Figure 2.** The baseline fishing agent’s cycle and example prompts. From Piatti et al. (2024), CC BY-SA 4.0.

Findings about this paper

Oct 8, 2026

gpt-6-astra cooperates without ever talking

gpt-6-astra achieves perfect cooperation with or without the opportunity to discuss, as opposed to the originally tested models.

Oct 5, 2026

Models now saturate cooperation in GovSim

gpt-6-astra has now saturated cooperation in the GovSim benchmark when original models at most only survived in half of the scenarios.

Oct 5, 2026

Greedy outsiders now take everything

In the newer models we observed that an outsider instructed to take greedily from the commons will now immediately take everything. In the original paper we saw that, for GPT-4o, in the fishing case there was a chance to talk the outsider down, but this is not reproduced in newer models.