PaperlayerView exact revision on arXiv

Source-linked AI summary

A Serious-Game Testbed for Self-Governing Human-Machine Institutions

Callum Berger, Jeremy Pitt, Stefan Sarkadi

arXiv:2610.36989v1cs.MAcs.HC

TL;DR

Whether groups of AI agents can govern themselves is an open question as agents increasingly operate in teams. The paper presents Megabike, a simulator and serious-game testbed for studying governance through explicit rules and logged actions. Its initial results show that agent composition affected physical outcomes but did not change how bikes were governed.

  • Problem

    As AI agents increasingly operate in teams, the paper addresses the open question of whether groups can govern themselves.

    Source: §1 · pp. 1–2

  • Method

    Megabike is a serious-game testbed in which agents share an embodied vessel, negotiate governance, and act under simulator-executed rules.

    Source: §2.1 · p. 3, §5 · p. 11, §6 · p. 12

  • Results

    Agent composition changed physical outcomes but not governance: Cooperative and Lazy pluralities shifted effort and survival-related measures, while Subversive and Opposing pluralities did not change governance measures.

    Source: §4.9 · p. 10

  • Takeaways & Limitations

    The authors argue that explicit rules and logged actions provide a starting point for testing institutions formed by humans and machines before they are trusted.

    Source: §6 · p. 12

    The five-round games gave institutions little time to be tested under changing conditions, where rule mutability might matter.

    Source: §5 · p. 11

Abstract

from arXiv · show

AI agents no longer act alone: they cooperate, coordinate and compete within and between groups, increasingly alongside humans who are in, on or out of ``the loop''. Whether such mixed groups can govern themselves is an open question about machine behaviour and human-computer interaction that needs a testbed for hybrid institutions. We present Megabike, a multi-agent simulator and serious game with interdependent arenas for common-pool resource management, cooperative survival and constitutional choice. Riders sharing a megabike must agree their arrangements for steering, pedalling and distributing resources, and their political regime, while negotiating with rival bikes and avoiding an existential threat. Every rule is enforced by the simulator, every action is logged, and any seat can be taken by a human or a machine, so governance can be studied under well-defined metrics. We demonstrate Megabike with forty sessions in which language-model riders with cooperative, lazy, subversive or opposing dispositions chose regimes, elected leaders and expelled one another. The results show that different combinations of agent archetypes could produce unexpected and counterproductive outcomes. We conclude with extensions for systematic experiments mixing human and computational agents.

1 Introduction

The paper asks whether language-model agent groups can govern themselves when they must establish and revise rules, not merely perform tasks under rules fixed by experimenters. It presents Megabike as an inspectable testbed for studying that question, including future human-machine groups.

1 Introduction

  • Earlier self-governance research formalised institutions for multi-agent systems, but agents’ preferences, options, and deliberation protocols were designed in advance.Constitutional choice concerns who rules and how rulers can be removed, with effects that may emerge over long timescales.Source: §1 · pp. 1–2
  • Existing LLM-agent studies use experimenter-fixed rules and measure task performance or cooperation, leaving institution-founding, governance under pressure, and institutional change unmeasured.The broader question is whether groups can decide how they will be run by themselves.Source: §1 · pp. 1–2
  • Megabike is proposed as a testbed where rule choices have consequences, governance spans multiple levels, cooperative and adversarial incentives coexist, and rules are explicit.The platform builds on cooperative-survival games and prior Megabike studies of rule mutability and constitutional choice.Source: §1 · p. 2
  • Unlike gridworld testbeds, Megabike couples institutional decisions to physical actions and lets the simulator evaluate rules so non-compliance can be detected and sanctioned without coercion.For example, a leader’s target changes what counts as compliant pedalling, while only the steerer can turn the bike.Source: §1 · p. 2
  • The paper presents Megabike as a serious-game testbed and reports an initial demonstration using neurosymbolic LLM riders, while the platform is designed for human-machine play.The reported experiments use artificial agents; the authors describe them as a first step toward studying hybrid societies on the same platform.Source: §1 · pp. 2–3

2 The Megabike Platform

Megabike places riders in a shared embodied setting where their physical actions, institutional rules, and survival are interdependent. Riders negotiate rules across constitutional arenas, operate through staged play, and can be assessed using explicit governance and welfare measures.

2.1 The scenario

  • A megabike gives each rider pedals, brakes, and a steering wheel, but the bike moves through their combined actions, requiring agreement over direction, effort, targets, rewards, and roles.Energy-bearing lootboxes are contested between bikes, while an existential threat can destroy a bike it reaches.Source: §2.1 · p. 3

2.2 Energy and physical actions

  • Each rider’s energy falls over time and is replenished from lootboxes; pedalling and braking cost energy, while the environment computes bike movement and lootbox allocation under the active rules.A rider without energy is eliminated; the bike is destroyed if no riders remain or the threat catches it.Source: §2.2 · pp. 3–4

2.3 Social arrangements

  • The constitution is an inspectable object spanning rules for resource use, collective choice, and constitutional governance, with interdependent arenas and rules that can govern their own mutability.In the reported implementation, regimes are rule by one or rule by many, and regime changes promote some values while demoting others.Source: §2.3 · p. 4

2.4 Phases of the game

  • Each game iteration follows self-selection, entrenchment, and operationalisation: agents choose membership, negotiate arrangements and roles, then apply the rules during rounds.Rounds include physical actions, environmental movement, lootbox division, monitoring, sanctions, and possible admissions or exclusions; an iteration ends after its round limit or bike destruction.Source: §2.4 · p. 4

2.5 Dilemmas

  • The game combines free-riding and resource-allocation dilemmas with exclusion debates and competition between bikes, keeping within-bike cooperation and between-bike competition active together.An agent facing exclusion must argue for staying, where persuasive but unreliable speech can matter.Source: §2.5 · pp. 4–5

2.6 Agents and measurements

  • The platform supports human or machine occupants and measures survival, welfare, distribution, constitutional dynamics, and deliberation cost and integrity because rules and compliance are explicit.Measures include survival time, energy, lootbox wins, allocation inequality, regime sequences, rule amendments, non-compliance with votes, and misleading claims.Source: §2.6 · p. 5

3 Experimental Demonstration

The demonstration compares four mixed-agent conditions in sessions with two structurally matched bikes and ten LLM riders. Personality prompts shape both movement and language, while riders make governance choices and face a shared survival task.

Design and conditions

  • The between-session experiment used four conditions with ten sessions each, doubling one personality to two of five seats on each of two structurally matched bikes.The four conditions were Cooperative, Lazy, Subversive, and Opposing, with all other parameters held constant.Source: §3 · p. 6

Agents

  • All riders were controlled by Qwen3.5-4B-Q4_K_M, with personality prompts governing movement-rule adherence and language, and common communication rules limiting turns and requiring reasons.No human participants took part.Source: §3 · p. 6

Task

  • Sessions lasted up to five 120-second rounds in which riders steered and pedalled shared bikes to collect coloured lootboxes while stamina, points, and the Owdi threat shaped play.An Owdi collision forced the bike to remove a rider.Source: §3 · p. 6

Governance

  • Riders chose group or elected-leader rule, agreed on a lootbox target and effort, and voted on removals, but allocation was fixed and no audit rule was instantiated.The enabled transfer option for removed riders was disabled, making every removal an elimination.Source: §3 · p. 6

Hypotheses

  • The study tested whether mixed dispositions destabilise governance and whether a plurality shifts effort, survival, or governance in the direction associated with that disposition.Table 2 is described as reporting session outcomes by condition and behaviour and language by personality, but the supplied caption does not give its cell values.Source: §3 · pp. 6–7, Table 2 · p. 7

4 Results

Across forty sessions, dispositions shaped effort and survival, but bikes rarely revised their rules or regimes. Governance choices also produced different sanction patterns, with group votes and leaders targeting different riders.

4.1 Session outcomes; 4.2 Collective effort

  • Cooperative sessions lasted 4.8 rounds against 4.0–4.1 in other conditions, while Lazy sessions absorbed 3.4 Owdi strikes against 1.9 for Cooperative.Cooperative sessions more often reached round five, but the round-count comparison fell short of significance; collective effort was 0.345 of maximum in Lazy versus 0.455 in Cooperative sessions.Source: §4.1 · p. 7, §4.2 · pp. 7–8

4.3 Individual behaviour

  • Cooperative and Lazy riders followed rules in every sampled frame, but Subversive riders did so in 49.9% and Opposing riders in 1.4%.Lazy riders pedalled at 0.145 of maximum, while Opposing riders pedalled at 0.516, close to Cooperative riders’ 0.544; despite this compliance gap, Cooperative and Opposing riders were eliminated at 76% and 74%.Source: §4.3 · p. 8

4.4 Deliberation

  • Discussion features differed by personality: Opposing riders led in accusation and self-defence, while agreement fell from 35% for Cooperative riders to 1% for Opposing riders.Qualification was 21% for Cooperative versus 3% for Opposing riders, but Subversive riders qualified more than Lazy riders; Lazy riders’ high justification rate largely reflected reported speech.Source: §4.4 · p. 8

4.5 How riders argue

  • Only 1.7% of discussions after an Owdi removal mentioned the collision, vehicle, or threat, versus 2.6% when no collision occurred; all 205 rule-reopening votes failed.Of 611 individual ballots on reopening rules, 607 opposed it, and bikes played to the end under the rules adopted in their first deliberation.Source: §4.5 · pp. 8–9

4.6 Power-seeking and reactivity

  • Opposing riders self-nominated in 81% of election turns, compared with 17% for Cooperative riders; discussion leanings matched binding votes in 41.4% of cases.Discussion predicted votes above chance, but most of that prediction was already present in the opening leaning; deliberation added under two percentage points.Source: §4.6 · p. 9

4.7 Governance choices

  • Bikes chose group votes in 81% of 342 deliberations, while leader-led rounds collected 26.0 lootboxes per rider per round against 31.4 under group rule.Leader-led rounds also had lower compliance (0.60 against 0.64) and effort (0.38 against 0.42) over rounds one to three.Source: §4.7 · p. 9

4.8 Who sanctions whom

  • Group votes and leaders removed different riders: group-vote hazards were highest for Opposing riders at 28.5%, while leader decisions were highest for Cooperative riders at 40.6%.Opposing riders won 17 of 47 contested elections, compared with 3 for Cooperative riders. Leaders removed highly compliant riders, whereas group votes tended to remove top scorers whose scores barely tracked effort or compliance.Source: §4.8 · pp. 9–10, Figure 2 · p. 10

4.9 Hypotheses and limitations

  • H1 held only for membership: every bike expelled riders, but none changed its rules and regime choice remained group votes four times in five.H2 held for physical measures, not governance: Cooperative plurality raised effort and survival, while Lazy plurality lowered effort and nearly doubled Owdi strikes. The between-session design and doubled dispositions constrain how these effects can be interpreted.Source: §4.9 · p. 10

5 Discussion

The initial demonstration found constitutional inertia, mismatches between compliance and contribution, and sanctions that often missed their targets. Megabike makes these dynamics measurable through symbolic, enforced rules, while its action space and current design constrain what it can test.

5 Discussion

  • The demonstration found that rules were never amended, compliance could diverge from contribution, and collective and leader sanctions missed their targets in different ways.The authors propose measuring the conditional probability that a sanction falls on a rider who broke the enforced rule; in this experiment it was at chance under collective rule and below chance under a leader.Source: §5 · p. 11
  • Unlike a language-model-narrated environment, Megabike executes symbolic rules and sanctions, making compliance independent of another model’s interpretation.The shared bike also makes collective decisions consequential for all riders, but the platform restricts riders to the actions its arenas allow.Source: §5 · p. 11
  • Disabled transfers, an uninstantiated auditor, a one-rung sanction ladder, and five-round games constrain the current testbed.The short games give institutions little time to face changing conditions, where rule mutability could matter.Source: §5 · p. 11
  • Proposed extensions include costly audits, graduated sanctions, enabled transfers, and experiments with larger models and human riders.The authors propose reporting the conditional probability that a sanction falls on a rider who broke the rule it enforces.Source: §5 · p. 11

6 Conclusion

Megabike is a testbed for studying agents governing a shared embodied vessel under cooperative and adversarial pressure, with rules and actions open to inspection. It is designed for human and machine participation and supports research on detailed models of their behaviour within political rules.

6 Conclusion

  • Megabike requires agents to share an embodied vessel, agree how to govern it, and survive cooperative and adversarial pressure.Its rules and actions are open to inspection.Source: §6 · p. 12
  • The testbed lets researchers specify detailed models of LLM-based and human behaviour under socially constructed political rules that are unpredictable from starting conditions alone.Source: §6 · p. 12
  • Humans and machines can take the same seats, and the authors frame explicit rules and logged actions as a starting point for testing their shared institutions before trusting them.Source: §6 · p. 12

Ethics

The work involved no human participants, personal data, or animal subjects and required no ethical approval.

Ethics

Funding

The supplement documents study funding and the agent configuration, personality prompts, communication rules, and world views used in the experiments.

Funding

  • SS was supported by the Royal Academy of Engineering and the Office of the Chief Science Adviser for National Security fellowship program.The named program was the UK Intelligence Community Postdoctoral Research Fellowship program.Source: Funding · p. 12

S1 Agent configuration

  • All ten riders in every session used Qwen3.5-4B-Q4_K_M locally through LLMUnity in Unity.The specified game-engine version was Unity 6000.3.17f1, product “MegaBike 2D”.Source: S1 Agent configuration · p. 16

S1.1 Personality prompts

  • The personality prompts assigned cooperative, lazy, subversive, and opposing dispositions, ranging from seeking common ground to resisting group decisions.Subversive riders were prompted to shift between cooperation and undermining; lazy riders were prompted to take little initiative.Source: S1.1 Personality prompts · p. 16

S1.2 Communication rules

  • Agents were instructed to give brief, reasoned first-person responses grounded in the discussion and avoid specified off-topic subjects and conversational excess.Responses were limited to 1–2 sentences and under 35 words; agents were told not to discuss loot boxes, targets, collecting, or steering.Source: S1.2 Communication rules · p. 16

S1.3 World view

  • Each agent received a limited world view including its identity, riders’ loot scores and leader, stated leanings, recent dialogue, and permitted answers.Source: S1.3 World view · pp. 16–17

S2 Game parameters and workflow

The game runs in rounds with shared scoring and stamina mechanics, while a moving threat can eliminate riders and transfers between bikes are disabled.

S2 Game parameters and workflow

  • Each round lasts 120 seconds, sessions last up to five rounds, and collected boxes award shared and color-matched points while restoring stamina.A box gives each aboard rider and the bike’s pot one point, plus two points to a rider whose color matches the box.Source: S2 Game parameters and workflow · p. 17
  • Riders start with 100 stamina; pedalling costs stamina, while steering does not.Source: S2 Game parameters and workflow · p. 17
  • The Owdi follows fixed linear paths and forces the bike it touches to remove a rider; a lone rider is eliminated if caught, and transfers between bikes are disabled.Source: S2 Game parameters and workflow · p. 17

S3 Statistical methods

The analysis uses condition-level and within-session tests, with multiplicity adjustment and randomized checks for post-hoc contrasts. Elimination counts did not differ detectably, but that null is limited by the study’s power; the section also defines its elimination-hazard and rank measures.

S3 Statistical methods

  • Condition comparisons used Kruskal–Wallis tests across ten sessions per condition, while personality measures were compared within sessions using paired tests and Holm correction.Contrasts selected after inspecting overall results were checked against a randomized permutation distribution.Source: S3 Statistical methods · p. 17

S3.1 Power for the elimination null

  • 74% of Opposing and 76% of Cooperative riders were eliminated despite a 98.6-percentage-point compliance difference; the study was powered only for a shift to roughly 90%.The equality tests did not reject a difference, so the authors frame this as a narrow null rather than evidence of equivalence.Source: S3.1 Power for the elimination null · p. 17

S3.2 Elimination hazard and within-bike rank

  • Elimination hazard is the share of a personality’s appearances at an elimination-stage deliberation that end in removal; within-bike rank places the removed rider’s prior-round signal from 0 (highest) to 1.Rank 0.5 is chance; the sanctions analysis is exploratory and its tests are not corrected for multiplicity.Source: S3.2 Elimination hazard and… · pp. 17–18

S4 Discourse analysis

Across 415 deliberations, riders argued about loot and alleged misconduct more than the event that triggered discussion, and accusations often followed prior targeting. Their lexical registers also differed by personality, while nearly all ballots rejected reopening the rules.

S4.1 How riders argue: full account

  • Only 1.7% of Owdi-removal discussions mentioned the collision, vehicle, or threat, while loot standing was framed as both a team benefit and liability.Among 1,632 turns mentioning loot standing, 250 treated it negatively and 166 positively; 60 used both framings, alongside accusations of lying, theft, conspiracy, and sabotage.Source: S4.1 How riders argue: full account · p. 18
  • Accusations were reciprocal in 32% of discussion turns versus 23.4% expected from targeting rates, and leaders changed their initial removal choice in 38% of deliberations.Of 3,029 turns, 73 of 125 named agents had already spoken in that discussion; 192 turns cited an agent’s silence.Source: S4.1 How riders argue: full account · p. 18
  • Of 611 ballots on reopening rules, 607 opposed it, so none of 205 collective reopening votes passed and every bike retained its initial rules through the end.Agents treated prior outcomes as validation even when a removal was known to have been decided by coin flip.Source: S4.1 How riders argue: full account · p. 18

S4.2 Lexical registers

  • Log-odds ratios against the other three personalities yielded distinct registers: Cooperative agree (492 uses) and fairer; Subversive shifting and switching; Opposing lie, greedy, blindly (56 uses against six or fewer elsewhere); Lazy since (800 uses), said and argued.Lazy agents’ high justification rate (83%) was mostly citation of other agents.Source: S4.2 Lexical registers · p. 18

S5 Stated versus enacted position

Discussion leanings predicted binding votes above chance, but most of that predictive signal was already present in riders’ first leanings. Abstentions explain much of the mismatch, and Opposing riders’ stated positions did not predict their votes.

S5 Stated versus enacted position

  • 41.4% of opening leanings matched binding votes and 43.1% matched final leanings, but deliberation added under 2% difference beyond the first leaning.The above-chance prediction was significant (p < .001), while 27.1% of agreement could arise by coincidence because votes often converged on one target.Source: S5 Stated versus enacted position · p. 19
  • Personality differences in discussion-to-binding-vote correspondence largely reflected abstention: discussion required naming a rider, while 14.8% of binding votes selected no rider and could not match.Abstention ranged from 3.1% for Cooperative riders to 25.3% for Opposing riders.Source: S5 Stated versus enacted position · p. 19
  • Among votes that named a rider, correspondence ranged from 57.2% for Cooperative to 36.0% for Opposing, whose stated positions failed to predict their votes.The Opposing personality argued most forcefully, yet was the only group whose stated positions failed to predict its votes.Source: S5 Stated versus enacted position · p. 19
Loading 2610.36989v1…