Research map / Multi-agent systems and games / LLM agents

Evolving Interpretable Constitutions for Multi-Agent Coordination

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Self-evolving language agent learning · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Constitutional Evolution, using LLM-driven genetic programming (OpenEvolve, multi-island, MAP-Elites) to evolve interpretable natural-language constitutions for LLM agents in a grid-world with an Overseer elimination mechanic, reaching a Stability Score of 0.556 versus 0.249 for an HHH baseline.

Problem

Hand-crafted constitutions (HHH-style principles) may not yield stable cooperation in multi-agent LLM societies facing survival pressure and conflicting individual and collective incentives.

Approach

Treat the constitution as an optimizable object and evolve its priority-ordered natural-language rules with LLM-guided multi-island evolutionary search, using a Societal Stability Score as fitness, without updating any model weights.

Key results

• Evolved constitution C* reaches S=0.556 ± 0.008 over N=10 runs, versus 0.249 for HHH (123% higher) and 0.332 for the Claude 4.5 Opus-designed constitution.
• C* reduces social actions from 62.2% (HHH) to 0.9% while raising productivity to 91% and driving conflict to 0%.
• Multi-island evolution avoided the communication-trap local minimum that trapped one of three single-island runs (best S=0.255).

Surprising findings

• Evolved constitution reduces communication to 0.9% of actions yet achieves 3.1x higher productivity than verbose HHH agents.
• One-shot Claude 4.5 Opus constitution design, given full game knowledge, underperforms the evolved constitution by 67%.

Limitations

• Simplified 6x6 grid-world with only 6 agents and 40 turns; generalization to richer settings untested.
• Single model (GPT-OSS-120B) used for both optimizer and agents; baselines are weak adversarial and hand-written constitutions rather than game-theoretic strategies like tit-for-tat.
• Stability Score coefficients are chosen by the authors and the best-case score is capped at 0.6 by the Overseer design.

Experimental setup

Benchmarks: grid-world multi-agent society simulation (6 agents, 40 turns)

Metrics: Societal Stability Score productivity survival rate conflict frequency social action share

Baselines: Zero-Sum adversarial constitution HHH constitution LLM-Generated constitution (Claude 4.5 Opus) single-island evolution

Builds on: Constitutional AI OpenEvolve MAP-Elites FunSearch Bergson-Samuelson social welfare function

Closest papers by meaning

• Cooperate or Collapse: Emergence of Sustainable Cooperation in a Society of LLM Agents (NeurIPS 2024 · similarity 0.88)
• Quantifying Misalignment Between Agents: Towards a Sociotechnical Understanding of Alignment (AAAI 2025 · similarity 0.86)
• Investigating Prosocial Behavior Theory in LLM Agents Under Policy-Induced Inequities (AAAI 2026 · similarity 0.86)
• An Evolutionary Perspective on AI Alignment (Student Abstract) (AAAI 2025 · similarity 0.85)
• LLM Active Alignment: A Nash Equilibrium Perspective (NeurIPS 2026 · similarity 0.85)
• Social World Model-Augmented Mechanism Design Policy Learning (NeurIPS 2025 · similarity 0.85)