Internal vs. External: Comparing Deliberation and Evolution for Multi-Agent Constitutional Design
A new academic study published on arXiv presents the first controlled comparison of internal deliberation and external evolution for designing behavioral constitutions in multi-agent AI systems. The research, conducted by Hershraj Niranjani and colleagues, evaluates these methods across three social environments: a coordination grid-world, an iterated public goods game, and a bilateral trading market. Based on 180 simulation runs, the findings indicate that external evolution significantly outperforms internal deliberation in collective-action settings. However, neither method yields improvements in bilateral trading scenarios. A critical ablation study reveals that external evolution's advantage can invert under shifted incentives, potentially forcing value-destroying cooperation. Notably, internal deliberation never proposed punishment mechanisms, whereas external optimization reliably discovered them. The authors conclude that while external optimization achieves higher performance peaks, internal self-governance offers greater structural responsiveness. This research addresses unresolved questions regarding whether AI governance rules should emerge from agent self-governance or be discovered through external optimization processes.
Wire timeline
Internal vs. External: Comparing Deliberation and Evolution for Multi-Agent Constitutional Design
A new academic study published on arXiv presents the first controlled comparison of internal deliberation and external evolution for designing behavioral constitutions in multi-agent AI systems. The research, conducted by Hershraj Niranjani and colleagues, evaluates these methods across three social environments: a coordination grid-world, an iterated public goods game, and a bilateral trading market. Based on 180 simulation runs, the findings indicate that external evolution significantly outperforms internal deliberation in collective-action settings. However, neither method yields improvements in bilateral trading scenarios. A critical ablation study reveals that external evolution's advantage can invert under shifted incentives, potentially forcing value-destroying cooperation. Notably, internal deliberation never proposed punishment mechanisms, whereas external optimization reliably discovered them. The authors conclude that while external optimization achieves higher performance peaks, internal self-governance offers greater structural responsiveness. This research addresses unresolved questions regarding whether AI governance rules should emerge from agent self-governance or be discovered through external optimization processes.
cs.AI updates on arXiv.org