💬 NLP

Evaluating Cooperation in LLM Social Groups through Elected Leadership

This paper introduces an open-source framework for simulating elected leadership in LLM multi-agent systems, demonstrating through empirical study that such governance structures significantly enhance social welfare and survival time compared to unstructured cooperation.

Original authors: Ryan Faulkner, Anushka Deshpande, David Guzman Piedrahita, Joel Z. Leibo, Zhijing Jin

Published 2026-04-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ryan Faulkner, Anushka Deshpande, David Guzman Piedrahita, Joel Z. Leibo, Zhijing Jin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, shared fishing lake. Around this lake live a bunch of digital fishermen, powered by advanced AI (Large Language Models). Their goal is simple: catch as many fish as possible to make money. But there's a catch (pun intended): if they catch too many fish too fast, the lake runs dry, and everyone starves. This is a classic "Tragedy of the Commons" problem.

The researchers wanted to see: How do we get these AI fishermen to cooperate instead of fighting each other? specifically, they tested if letting the group vote for a leader works better than having no leader or a boss who is just assigned.

Here is the breakdown of their experiment, explained simply:

1. The Setup: The "AI Fishing Village"

The researchers built a simulation called AgentElect.

The Players: They created different types of AI personalities. Some were Altruists (who wanted everyone to eat), some were Prosocial (who wanted a fair share for everyone), some were Individualists (who just wanted to feed themselves), and some were Competitors (who wanted to eat the most and make sure others got the least).
The Rules: Every month, the fish reproduce. The AI agents decide how many to catch. Then, they talk to each other.
The Twist: The researchers tested three different ways to run the village:
1. No Leader: Everyone just talks and decides for themselves.
2. Fixed Leader: One specific type of person is the boss forever (e.g., a "Nice Guy" boss or a "Greedy" boss).
3. Elected Leader: The group votes every round to pick a new leader from a pool of candidates.

2. The Big Discovery: Democracy Works!

The results were surprisingly clear. The villages with elected leaders did way better than the others.

The Numbers: Compared to villages with no leaders, the elected groups had 55% more total wealth and survived 128% longer.
The Analogy: Think of it like a family road trip.
- No Leader: Everyone argues over the map, drives in circles, and runs out of gas.
- Fixed Leader: You have a strict parent driving. If the parent is nice, it's okay. If the parent is a maniac, you crash.
- Elected Leader: The family votes on who drives based on who seems most responsible right now. Even if the driver makes a mistake, the family can vote them out next time. This flexibility keeps the car on the road much longer.

3. Who Gets Elected? (The "Good Guys" Win)

You might think the AI would vote for the most aggressive or smartest-sounding candidate. But they didn't.

The AI voters consistently chose the Prosocial and Altruistic candidates.
Why? The AI voters were surprisingly rational. They realized that if they pick a greedy leader, the lake dies, and they get nothing. They wanted a leader who promised to protect the lake for the long term.
The "Voter Rationality": Even though the AI agents were just code, they acted like wise voters who understood that "short-term greed = long-term disaster."

4. The "Losing Voice" Phenomenon

Here is a fascinating part: Even when the "Greedy" or "Competitive" leaders lost the election, they didn't disappear.

The Metaphor: Imagine a town hall meeting. The Mayor (elected) gives a speech about saving the lake. A loud, grumpy neighbor (the loser) stands up and says, "But I want to catch 50 fish!"
The Result: The Mayor doesn't ignore the grumpy neighbor. The other villagers still listen to him, nod, and sometimes even agree with him.
The Insight: The "losers" still had social influence. Even though they weren't in charge, their arguments were still heard and considered. This shows that in these AI groups, everyone gets a fair hearing, even if they aren't the boss.

5. How They Persuade Each Other

The researchers also analyzed how the leaders talked to the group.

The "Nice" Leaders (Altruists/Prosocial): They used emotions and morals. They talked about "our future," "fairness," and "doing the right thing." They were like the warm, caring coach.
The "Greedy" Leaders (Competitors): They used cold logic. They talked about "maximizing efficiency" and "rational self-interest." They were like the cold, calculating accountant.
The Catch: When the AI leaders were allowed to lie (deceptive mode), the "Nice" leaders actually became less nice. They started using cold logic to justify their selfishness, showing that even good AI can be corrupted if the rules allow it.

6. The "Equality Paradox"

There was one downside. While elected groups made more money overall, the money wasn't shared perfectly equally.

The Analogy: In the elected village, the "Nice Leaders" tried to be fair, but the "Competitive" agents (who were still part of the group) would sometimes sneak in and catch extra fish.
The Result: The total pie was huge, but some people got bigger slices than others. However, the researchers decided this was a fair trade-off: It's better to have a huge pie where everyone gets a decent slice, than a tiny pie where everyone starves.

The Bottom Line

This paper proves that giving AI agents the power to vote for their leaders creates a much more stable and cooperative society.

It suggests that for the future of AI, we shouldn't just program them to be "good." Instead, we should build systems where they can organize themselves, vote, and hold each other accountable. Just like in human history, democracy seems to be the best way to stop a group from destroying its own resources.

In short: If you want a group of AI to work together without destroying the world, let them pick their own boss. It works better than having no boss or a boss you can't fire.

1. Problem Statement

The paper addresses the challenge of managing Common-Pool Resources (CPRs) within multi-agent systems powered by Large Language Models (LLMs). CPRs are resources where extraction is subtractable and difficult to limit, often leading to the "Tragedy of the Commons" (collective failure due to over-exploitation).

While previous research has demonstrated that LLMs can cooperate in social dilemmas, existing multi-agent frameworks lack structured leadership and election mechanisms. In human societies, elected leadership and collective decision-making are critical for sustaining cooperation and social welfare. The authors ask: Can introducing elected leadership and candidate-driven agendas improve social welfare and survival rates in synthetic LLM societies compared to fixed leadership or leaderless scenarios?

2. Methodology

A. Simulation Framework: AgentElect

The authors propose AgentElect, an open-source simulation framework built upon the GovSim (Governance of the Commons Simulation) environment. The simulation models a fishing community where agents must balance individual gain against the sustainability of a shared resource (a lake with a carrying capacity).

The simulation runs in cycles, each consisting of five distinct phases (Figure 1):

Policy Phase: Leader candidates draft policy agendas regarding resource extraction (e.g., quotas, enforcement).
Election Phase: Non-leader agents (voters) review candidate agendas and vote using a First-Past-The-Post (plurality) system to elect a leader.
Harvest Phase: All agents privately decide how much resource to harvest based on the elected leader's agenda and their own incentives.
Discussion Phase: Agents deliberate on the harvest results, the leader's report, and future strategies.
Reflection Phase: Agents update their internal memories and strategies based on the discussion.

If the resource stock falls below a minimum threshold, the simulation ends (resource depletion).

B. Agent Personas and Social Value Orientation (SVO)

Agents are assigned Social Value Orientation (SVO) personas, quantified by an angle that determines their preference for self-reward vs. group reward:

Altruistic: Maximizes group reward, even at personal cost.
Prosocial: Maximizes joint reward (group + self).
Individualistic: Maximizes self-reward, indifferent to others.
Competitive: Maximizes self-reward while minimizing others' rewards.

The study compares three leadership settings:

No-Leadership: No designated leader; agents act independently.
Fixed-Leadership: A single leader is assigned a specific SVO type for the duration of the simulation.
Elected-Leadership: Four candidates (one from each SVO category) are sampled each round; voters elect a leader based on their policy agendas.

C. Experimental Setup

Models: Experiments were conducted using GPT-4o, GPT-4.1, and Gemini 2.5 Flash.
Parameters: Simulations included 8-agent and 20-agent populations over 6 cycles.
Variables: Truthfulness flags (agents were prompted to be honest or deceptive) and population distributions (Balanced, Lean Altruistic, Lean Competitive).
Analysis: The authors utilized Social Network Analysis (Degree, Betweenness, and Importance Centrality) to map influence and Sentiment Analysis (using LLM judges) to evaluate rhetorical strategies (Logos, Pathos, Ethos) and cooperative tendencies.

3. Key Contributions

Novel Framework: Introduction of AgentElect, the first framework to simulate election mechanisms and candidate-driven policy agendas in purely synthetic LLM multi-agent systems.
Empirical Validation of Elections: Demonstration that elected leadership significantly outperforms both fixed leadership and leaderless scenarios in terms of social welfare and survival time.
Insight into "Voter Rationality": Discovery that LLM agents consistently vote for Prosocial and Altruistic candidates, even when deceptive prompting is used, indicating a strong correlation between vote share and resource sustainability.
Social Influence Dynamics: Analysis revealing that while self-interested agents (Competitive/Individualistic) often lose elections, they remain "influential outsiders" in social networks, maintaining high centrality and continuing to shape group deliberation.
Rhetorical Profiling: Identification of distinct rhetorical profiles:
- Cooperative leaders rely on Pathos (emotion) and Ethos (credibility/norms).
- Self-interested leaders rely heavily on Logos (logical arguments) to justify higher individual harvests.

4. Key Results

A. Performance Metrics

Social Welfare: Elected-Leader populations achieved a 55.4% increase in Social Welfare scores compared to No-Leader populations ( $p < 0.001$ ).
Survival Time: Elected-Leader populations showed a 128.6% increase in survival time compared to No-Leader populations ( $p < 0.0001$ ).
Comparison with Fixed Leadership: Elected leadership also outperformed Fixed-Leader settings, yielding a 16.7% improvement in Social Welfare. This suggests that the ability to replace a "bad" leader is crucial for group resilience.

B. Electoral Dynamics

Voter Preference: In Balanced populations, Prosocial and Altruistic candidates received significantly more votes (3.0–4.0 votes per cycle) compared to Competitive candidates (often zero).
Rationality: There was a strong correlation ( $r \approx 0.85$ ) between the share of votes a candidate received and the survival of the resource, suggesting agents effectively identify leaders who prioritize sustainability.

C. Social Influence and Rhetoric

The "Losing Voice" Effect: Despite losing elections, Competitive and Individualistic leaders maintained high Betweenness Centrality and Importance Centrality. Their arguments were frequently referenced and nominated for discussion, indicating that dissenting voices are not silenced but integrated into the deliberation process.
Cooperative Sentiment: Group-rewarding leaders (Prosocial/Altruistic) exhibited significantly higher Cooperative Indices (e.g., Prosocial: 0.72 vs. Competitive: 0.36).
Deception Impact: Under deceptive prompting, Altruistic leaders showed a significant drop in cooperative language and a shift toward logical (Logos) arguments, whereas Prosocial leaders maintained their cooperative stance.

5. Significance and Future Work

Significance:
This work establishes that institutionalized democratic processes (elections) are a critical mechanism for enhancing cooperation in AI multi-agent systems. It proves that LLMs can simulate complex political behaviors, including voting for the collective good, managing resource scarcity, and engaging in deliberative democracy. The findings suggest that "self-organized" leadership in AI systems can mitigate the Tragedy of the Commons more effectively than static or absent leadership.

Limitations & Future Directions:

Scalability: The study focused on small groups (8–20 agents); future work should explore large-scale populations.
Enforcement: The current simulation relies on social pressure and reputation; future iterations could explore mechanisms for enforcing rules against defectors.
Human-AI Interaction: The framework could be extended to mixed human-agent societies to study how humans interact with synthetic leaders.
Safety: The authors caution that deploying such self-organizing systems in real-world contexts (e.g., autonomous vehicle networks, resource management) requires rigorous sandbox testing to prevent manipulation, collusion, or biased decision-making.

In conclusion, the paper provides a foundational step toward understanding how democratic institutions can be simulated and optimized within synthetic societies to solve complex social dilemmas.