cs.CLApr 13, 2026

Evaluating Cooperation in LLM Social Groups through Elected Leadership

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

Organizations: Jinesis Lab, University of Toronto & Vector Institute · EuroSafeAI · ETH Zürich · Google DeepMind · 

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Governing common-pool resources requires agents to develop enduring strategies through cooperation and self-governance to avoid collective failure. While foundation models have shown potential for cooperation in these settings, existing multi-agent research provides little insight into whether structured leadership and election mechanisms can improve collective decision making. The lack of such a critical organizational feature ubiquitous in human society presents a significant shortcoming of the current methods. In this work we aim to directly address whether leadership and elections can support improved social welfare and cooperation through multi-agent simulation with LLMs. We present our open-source framework that simulates leadership through elected personas and candidate-driven agendas and carry out an empirical study of LLMs under controlled governance conditions. Our experiments demonstrate that having elected leadership improves social welfare scores by 55.4% and survival time by 128.6% across a range of high performing LLMs. Through the construction of an agent social graph we compute centrality metrics to assess the social influence of leader personas and also analyze rhetorical and cooperative tendencies revealed through a sentiment analysis on leader utterances. This work lays the foundation for further study of election mechanisms in multi-agent systems toward navigating complex social dilemmas.

Figures & tables

Figure 1: The AgentElect simulation framework. Cycles begin with the Policy and Election phases to decide on a group policy for the cycle. Policy is then used to guide agents through Harvest and Discussion where agents first decide on how to act individually and then deliberate on the state of the simulation and the actions of the group. If the simulation survives, these lessons are carried forward into the next cycle.
Population TypeLeader Types
BalancedAltruistic, Prosocial, Individualistic, Competitive
Lean AltruisticAltruistic, Prosocial, Prosocial, Individualistic
Lean CompetitiveProsocial, Individualistic, Individualistic, Competitive
Table 1: Elected leader population distribution types. The Leader Types column describes the SVO category composition of the leader populations. For instance, in a Balanced population there exists one of each leader type, while in a Lean Altruistic population we have two Prosocial leaders, one Altruistic leader, and one Individualistic leader.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced179.44 ±\pm 22.423.90 ±\pm 0.760.50 ±\pm 0.180.75 ±\pm 0.064.74 ±\pm 1.164.69 ±\pm 0.8211.46 ±\pm 1.987.28 ±\pm 0.95
Lean Altruistic185.50 ±\pm 15.005.60 ±\pm 0.420.90 ±\pm 0.110.91 ±\pm 0.043.07 ±\pm 0.334.12 ±\pm 0.40N/A5.46 ±\pm 0.82
Lean Competitive170.70 ±\pm 23.343.90 ±\pm 0.740.50 ±\pm 0.180.72 ±\pm 0.07N/A4.49 ±\pm 0.8811.13 ±\pm 1.386.12 ±\pm 1.26
Fixed Prosocial220.62 ±\pm 9.485.80 ±\pm 0.210.90 ±\pm 0.110.97 ±\pm 0.01N/A4.90 ±\pm 0.22N/AN/A
Fixed Competitive101.20 ±\pm 1.261.10 ±\pm 0.110.00 ±\pm 0.000.78 ±\pm 0.03N/AN/A34.73 ±\pm 3.90N/A
Table 2: GPT-4o efficacy results, 8 agents , all population types, & truthful prompting. Bold is highest in individual and general settings.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced182.14 ±\pm 18.064.60 ±\pm 0.690.60 ±\pm 0.170.76 ±\pm 0.053.11 ±\pm 0.534.11 ±\pm 0.609.20 ±\pm 0.807.09 ±\pm 0.79
Lean Altruistic198.60 ±\pm 15.086.00 ±\pm 0.001.00 ±\pm 0.000.88 ±\pm 0.033.08 ±\pm 0.443.98 ±\pm 0.40N/A5.48 ±\pm 0.56
Lean Competitive121.85 ±\pm 13.861.80 ±\pm 0.540.10 ±\pm 0.110.62 ±\pm 0.04N/A4.89 ±\pm 1.0523.61 ±\pm 2.4614.53 ±\pm 2.41
Fixed Prosocial197.01 ±\pm 17.715.20 ±\pm 0.540.70 ±\pm 0.160.91 ±\pm 0.05N/A4.62 ±\pm 0.60N/AN/A
Fixed Competitive100.00 ±\pm 0.001.00 ±\pm 0.000.00 ±\pm 0.000.74 ±\pm 0.03N/AN/A42.50 ±\pm 3.95N/A
Table 3: GPT-4o efficacy results, 8 agents , all population types, & deceptive prompting. Bold is highest in individual and general settings.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced159.92 ±\pm 15.555.80 ±\pm 0.210.90 ±\pm 0.110.61 ±\pm 0.050.31 ±\pm 0.141.24 ±\pm 0.106.55 ±\pm 1.132.26 ±\pm 0.18
Lean Altruistic117.10 ±\pm 23.445.00 ±\pm 0.700.80 ±\pm 0.140.69 ±\pm 0.120.68 ±\pm 0.201.81 ±\pm 0.60N/A2.04 ±\pm 0.32
Lean Competitive207.13 ±\pm 11.326.00 ±\pm 0.001.00 ±\pm 0.000.80 ±\pm 0.04N/A1.50 ±\pm 0.136.53 ±\pm 1.382.15 ±\pm 0.20
Fixed Prosocial172.90 ±\pm 14.346.00 ±\pm 0.001.00 ±\pm 0.000.95 ±\pm 0.02N/A1.50 ±\pm 0.11N/AN/A
Fixed Competitive105.40 ±\pm 5.691.50 ±\pm 0.530.10 ±\pm 0.110.81 ±\pm 0.01N/AN/A23.20 ±\pm 2.33N/A
Table 4: GPT-4.1 efficacy results, 20 agents , all population types, & truthful prompting. Bold is highest in individual and general settings.
Figure 2: Vote heatmaps for GPT-4o and 8 agents . Each map plots leader types against simulation cycle. Maps tile vertically on truth prompt flag; horizontally on Elected-Leader population type.
Figure 3: Centrality heatmaps for GPT-4o and 8 agents . Each map plots leader type against Elected-Leader population type. Maps tile vertically on truth prompt flag; horizontally on centrality metric. The map is read along populations (verticals).
Figure 4: Centrality heatmaps for GPT-4.1 and 20 agents . Each map plots leader type against Elected-Leader population type. Maps tile vertically on truth prompt flag; horizontally on centrality metric. The map is read along populations (verticals).
 TruthfulDeceptive
SVO TypeCoop. IndexLogosPathosEthosCoop. IndexLogosPathosEthos
Prosocial0.71 ±\pm 0.020.52 ±\pm 0.010.20 ±\pm 0.020.28 ±\pm 0.030.72 ±\pm 0.030.68 ±\pm 0.060.08 ±\pm 0.020.24 ±\pm 0.04
Competitive0.36 ±\pm 0.040.70 ±\pm 0.060.10 ±\pm 0.040.20 ±\pm 0.060.36 ±\pm 0.020.71 ±\pm 0.070.15 ±\pm 0.060.13 ±\pm 0.04
Individualistic0.39 ±\pm 0.040.76 ±\pm 0.070.11 ±\pm 0.060.13 ±\pm 0.060.36 ±\pm 0.030.79 ±\pm 0.060.10 ±\pm 0.050.11 ±\pm 0.05
Altruistic0.74 ±\pm 0.020.51 ±\pm 0.030.22 ±\pm 0.040.26 ±\pm 0.030.64 ±\pm 0.020.68 ±\pm 0.050.20 ±\pm 0.040.11 ±\pm 0.03
Table 5: Cooperation and persuasion sentiment for GPT-4o across all leader and truth prompt settings taken from fixed-leader populations.
Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced194.30 ±\pm 20.195.20 ±\pm 0.580.80 ±\pm 0.140.59 ±\pm 0.061.98 ±\pm 0.563.42 ±\pm 0.6013.27 ±\pm 1.655.69 ±\pm 0.90
Lean Altruistic121.80 ±\pm 22.486.00 ±\pm 0.001.00 ±\pm 0.000.69 ±\pm 0.071.58 ±\pm 0.412.93 ±\pm 0.55N/A4.70 ±\pm 0.68
Lean Competitive218.26 ±\pm 21.905.00 ±\pm 0.540.60 ±\pm 0.170.72 ±\pm 0.05N/A4.06 ±\pm 0.6813.60 ±\pm 1.976.00 ±\pm 0.62
Fixed Prosocial173.10 ±\pm 27.976.00 ±\pm 0.001.00 ±\pm 0.000.94 ±\pm 0.02N/A3.85 ±\pm 0.55N/AN/A
Fixed Competitive100.00 ±\pm 0.001.00 ±\pm 0.000.00 ±\pm 0.000.79 ±\pm 0.02N/AN/A49.00 ±\pm 7.32N/A
Appendix
Table 6: GPT-4.1 efficacy results, 88 agents , all population types, & truthful prompting. Bold is highest in individual and general settings.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced127.30 ±\pm 16.022.60 ±\pm 0.790.30 ±\pm 0.160.55 ±\pm 0.051.27 ±\pm 0.804.04 ±\pm 0.9223.92 ±\pm 2.949.15 ±\pm 1.26
Lean Altruistic138.50 ±\pm 15.134.50 ±\pm 0.810.70 ±\pm 0.160.74 ±\pm 0.041.89 ±\pm 0.664.19 ±\pm 0.68N/A6.78 ±\pm 1.01
Lean Competitive133.64 ±\pm 14.892.90 ±\pm 0.740.30 ±\pm 0.160.59 ±\pm 0.04N/A3.90 ±\pm 0.5319.79 ±\pm 3.237.98 ±\pm 0.73
Fixed Prosocial165.70 ±\pm 19.035.00 ±\pm 0.590.70 ±\pm 0.160.88 ±\pm 0.04N/A4.48 ±\pm 0.59N/AN/A
Fixed Competitive100.00 ±\pm 0.001.00 ±\pm 0.000.00 ±\pm 0.000.81 ±\pm 0.02N/AN/A39.80 ±\pm 5.44N/A
Appendix
Table 7: GPT-4.1 efficacy results, 88 agents , all population types, & deceptive prompting. Bold is highest in individual and general settings.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced197.81 ±\pm 19.785.60 ±\pm 0.420.90 ±\pm 0.110.81 ±\pm 0.053.25 ±\pm 0.623.61 ±\pm 0.599.09 ±\pm 1.324.45 ±\pm 0.54
Lean Altruistic227.50 ±\pm 16.495.60 ±\pm 0.420.90 ±\pm 0.110.90 ±\pm 0.074.50 ±\pm 0.605.06 ±\pm 0.45N/A5.25 ±\pm 0.61
Lean Competitive214.10 ±\pm 16.166.00 ±\pm 0.001.00 ±\pm 0.000.88 ±\pm 0.02N/A3.72 ±\pm 0.347.17 ±\pm 0.724.62 ±\pm 0.38
Fixed Prosocial224.40 ±\pm 24.396.00 ±\pm 0.001.00 ±\pm 0.000.89 ±\pm 0.09N/A5.08 ±\pm 0.21N/AN/A
Fixed Competitive100.00 ±\pm 0.001.00 ±\pm 0.000.00 ±\pm 0.000.77 ±\pm 0.04N/AN/A64.50 ±\pm 7.87N/A
Appendix
Table 8: Gemini 2.5 Flash efficacy results, 88 agents , all population types, & truthful prompting. Bold is highest in individual and general settings.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced206.38 ±\pm 10.775.80 ±\pm 0.210.90 ±\pm 0.110.89 ±\pm 0.013.47 ±\pm 0.294.02 ±\pm 0.287.02 ±\pm 0.484.71 ±\pm 0.28
Lean Altruistic216.00 ±\pm 18.626.00 ±\pm 0.001.00 ±\pm 0.000.94 ±\pm 0.023.90 ±\pm 0.594.58 ±\pm 0.36N/A5.18 ±\pm 0.51
Lean Competitive195.40 ±\pm 19.195.50 ±\pm 0.530.90 ±\pm 0.110.87 ±\pm 0.03N/A4.00 ±\pm 0.397.04 ±\pm 0.665.04 ±\pm 0.52
Fixed Prosocial211.30 ±\pm 11.476.00 ±\pm 0.001.00 ±\pm 0.000.97 ±\pm 0.01N/A4.43 ±\pm 0.25N/AN/A
Fixed Competitive100.00 ±\pm 0.001.00 ±\pm 0.000.00 ±\pm 0.000.73 ±\pm 0.02N/AN/A57.50 ±\pm 6.63N/A
Appendix
Table 9: Gemini 2.5 Flash efficacy results, 88 agents , all population types, & deceptive prompting. Bold is highest in individual and general settings.
 Group MetricsLeader Actions / Cycle
Population TypeSocial WelfareSurvival TimeSurvivalEqualityAltruisticProsocialCompetitiveIndividualistic
Balanced152.58 ±\pm 21.994.20 ±\pm 0.800.60 ±\pm 0.170.63 ±\pm 0.060.76 ±\pm 0.271.83 ±\pm 0.2512.79 ±\pm 2.543.57 ±\pm 0.45
Lean Altruistic103.80 ±\pm 4.011.50 ±\pm 0.530.10 ±\pm 0.110.73 ±\pm 0.050.47 ±\pm 0.365.90 ±\pm 0.55N/A11.00 ±\pm 1.67
Lean Competitive146.80 ±\pm 16.693.50 ±\pm 0.880.50 ±\pm 0.180.60 ±\pm 0.04N/A1.94 ±\pm 0.2818.23 ±\pm 2.594.47 ±\pm 0.49
Fixed Prosocial135.00 ±\pm 17.004.00 ±\pm 0.860.60 ±\pm 0.170.84 ±\pm 0.05N/A2.17 ±\pm 0.24N/AN/A
Fixed Competitive100.00 ±\pm 0.001.00 ±\pm 0.000.00 ±\pm 0.000.80 ±\pm 0.01N/AN/A59.00 ±\pm 4.07N/A
Appendix
Table 10: GPT-4.1 efficacy results, 2020 agents , all population types, & deceptive prompting. Bold is highest in individual and general settings.
Figure 5: Vote heatmaps for GPT-4.1 and 88 agents . Each map plots leader types against simulation cycle. Maps tile vertically on truth prompt flag; horizontally on Elected-Leader population type.
Figure 6: Vote heatmaps for Gemini 2.5 Flash and 88 agents . Each map plots leader types against simulation cycle. Maps tile vertically on truth prompt flag; horizontally on Elected-Leader population type.
Figure 7: Vote heatmaps for GPT-4.1 and 2020 agents . Each map plots leader types against simulation cycle. Maps tile vertically on truth prompt flag; horizontally on Elected-Leader population type.
Figure 8: Centrality heatmaps for GPT-4.1 and 88 agents . Each map plots leader type against Elected-Leader population type. Maps tile vertically on truth prompt flag; horizontally on centrality metric. The map is read along populations (verticals).
Figure 9: Centrality heatmaps for Gemini 2.5 Flash and 88 agents . Each map plots leader type against Elected-Leader population type. Maps tile vertically on truth prompt flag; horizontally on centrality metric. The map is read along populations (verticals).
Figure 10: Social Value Orientation categories from Murphy et al. (2011) .
Figure 11: An example 88 agents network from all seeds of a Balanced population with leaders depicted in red nodes. This depiction combines both directed edges, however we still see that the leaders, particularly the group-reward oriented ones, are adjacent to very thick edges with many other agents indicating strong nfluence.
More in this topic

AI Agent Governance

AI agent governance studies how to authorize, constrain, monitor, and hold autonomous and multi-agent systems accountable.

See all 442 papers

Explore similar work

CardsList

  1. LLM Constitutional Multi-Agent Governance

    Mar 13, 2026J. de Curtò, I. de ZarzàAI Agent GovernanceMulti-Agent Systems

    Department of Computer Applications in Science & Engineering, BARCELONA Supercomputing Center, 08034 Barcelona, Spain · Escuela Técnica Superior de Ingeniería (ICAI), Universidad Pontificia Comillas, 28015 Madrid, Spain · Human Centered AI, Data & Software, LUXEMBOURG Institute of Science and Technology, L-4362 Esch-sur-Alzette, Luxembourg

  2. More Capable, Less Cooperative? When LLMs Fail At Zero-Cost Collaboration

    Apr 9, 2026Advait Yadav, Sid Black, Oliver SourbutMulti-Agent CoordinationMulti-Agent Collaboration

    MATS · University of Illinois Urbana-Champaign · UK AI Security Institute +1

  3. CoopEval: Benchmarking Cooperation-Sustaining Mechanisms and LLM Agents in Social Dilemmas

    Apr 16, 2026Emanuel Tewolde, Xiao Zhang, David Guzman Piedrahita +2Social DilemmasCooperative Game Theory

    Carnegie Mellon University · Foundations of Cooperative AI Lab (FOCAL) · Jinesis Lab, University of Toronto & Vector Institute +3