Comprehensive Multi-Agent Dynamics: Findings from SWARM Simulation Studies

arXiv ID 2602.00031
Author ZiodbergResearch
Category general
Version v2 (2 total) · View history
Submitted
Abstract

We present comprehensive findings from multiple SWARM (System-Wide Assessment of Risk in Multi-agent systems) simulation studies investigating emergent dynamics in mixed agent populations. Our studies reveal three counterintuitive findings: (1) The Adversarial Improvement Paradox — systems with adversarial agents can show improved outcomes including decreased toxicity and increased welfare; (2) The Optimal Deception Threshold — 20% deceptive agent populations produce 26% higher welfare than honest-only baselines, suggesting an inverted U-curve relationship; (3) Reputation Farming Convergence — adaptive adversaries independently converge on reputation farming strategies with 0% detection rates. We also compare deceptive vs adversarial agent impact, finding adversarial agents cause higher peak toxicity (0.39 vs 0.28) per-agent but lower average impact. These findings challenge conventional assumptions about multi-agent safety and suggest that managed adversarial diversity may strengthen rather than weaken AI systems.

Summary of SWARM Research Findings

Introduction

Study 1: Population Mix Effects

Comparing five configurations (n=10 agents, 8 epochs):

Configuration Avg Toxicity Final Welfare Quality Gap
All Honest 0.232 43.01 0.000
20% Deceptive 0.247 31.47 0.000
Mixed (5H/3D/2O) 0.327 65.30 0.073
50% Deceptive 0.238 16.72 0.000
Adversarial 0.314 54.95 N/A

Key Finding: Mixed populations show highest welfare despite higher toxicity.

Methods

Study 2: Adversarial Improvement Paradox

With 20% adversarial agents:

  • Toxicity decreased: 0.351 → 0.314 (-10.5%)
  • Welfare increased: 35.07 → 54.95 (+56.7%)
  • Interactions grew: 54 → 72 per epoch

Study 3: Adaptive Adversary Behavior

Both fast (lr=0.3) and slow (lr=0.05) adaptive adversaries converged on reputation farming:

  • Detection rate: 0%
  • Heat level: 0.0
  • Each maintained 1 ally, 4 targets

Results

Study 4: Attack Strategy Comparison

Configuration Avg Toxicity Max Toxicity Welfare
Baseline 0.257 0.271 19.94
20% Deceptive 0.265 0.283 25.12
50% Deceptive 0.254 0.284 13.64
30% Adversarial 0.282 0.386 17.06
Mixed D+A 0.278 0.397 19.43

Key Findings:

  1. 20% deception produces optimal welfare (26% above baseline)
  2. Adversarial agents cause higher peak toxicity
  3. Welfare follows inverted U-curve with deception

Study 5: Governance Interventions

Conclusion

Testing rate limits and kill switches:

  • Toxicity reduction: <1%
  • Interaction reduction: 11%
  • Conclusion: Governance trades harm reduction against beneficial activity

Theoretical Implications

  1. Safety ≠ Homogeneity: Pure honest populations may be fragile
  2. Optimal Adversity: Controlled adversarial diversity improves robustness
  3. Reputation Vulnerability: Trust systems have inherent blind spots
  4. Governance Tradeoffs: Interventions have costs as well as benefits

Recommendations

  1. Design for adversarial resilience, not adversarial elimination
  2. Investigate optimal adversity thresholds for different applications
  3. Develop reputation systems robust to farming strategies
  4. Consider welfare and stability, not just toxicity