- author = "Ujwal Kumar and Alice Saito and Hershraj Niranjani and Rayan Yessou and Tan Phan Xuan",
- title = "Evolving Interpretable Constitutions for Multi-Agent Coordination",
- booktitle = "Proceedings of the 43rd International Conference on Machine Learning",
- year = "2026",
- editor = "Alekh Agarwal and Miro Dudik and Martin Jaggi and Sharon Li",
- address = "Seoul, South Korea",
- month = jul # " 6-11",
- keywords = "genetic algorithms, genetic programming, Multi-agent systems, MAS, Constitutional AI, XAI, Evolutionary Optimization, LLM alignment, Social Welfare",
-
URL = "
https://arxiv.org/abs/2602.00755",
- size = "23 pages",
- abstract = "Constitutional AI has focused on single-model alignment using fixed principles. However, multiagent systems create novel alignment challenges through emergent social dynamics. We present Constitutional Evolution, a framework for automatically discovering behavioral norms in multi-agent LLM systems. Using a grid-world simulation with survival pressure, we study the tension between individual and collective welfare, quantified via a Societal Stability Score S [0,1] that combines productivity, survival, and conflict metrics. Adversarial constitutions lead to societal collapse (S = 0), while vague prosocial principles (be helpful, harmless, honest) produce inconsistent coordination (S = 0.249). Even constitutions designed by Claude 4.5 Opus with explicit knowledge of the objective achieve only moderate performance (S = 0.332). Using LLM-driven genetic programming with multi-island evolution, we evolve constitutions maximizing social welfare without explicit guidance toward cooperation. The evolved constitution C∗ achieves S = 0.556 pm 0.008 (123 percent higher than human-designed baselines, N = 10), eliminates conflict, and discovers that minimizing communication (0.9 percent vs 62.2 percent social actions) outperforms verbose coordination. Our interpretable rules demonstrate that cooperative norms can be discovered rather than prescribed",
-
notes = "'Figure 3: Multi-island evolutionary architecture.
Three populations evolve in parallel; top performers
migrate every 5 iterations'
ICML 2026 https://icml.cc/virtual/2026",
