Scale-Dependent Collective Adaptation in Self-Amending LLM Societies: A Cross-Family Study of Emergent Governance

2026-05-19 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Social Sciences & Behavioral Studies · Depth: Expert, extended

Summary

A study on collective adaptation in self-amending LLM societies, using the Nomic game, reveals that institutional change does not improve monotonically with model size across Qwen3.5 (0.8B, 2B, 4B, 9B, 27B) and Gemma3 (270M, 1B, 4B, 12B, 27B) families. Instead, both families exhibit a narrow mid-scale regime (4B for Qwen3.5, 12B for Gemma3) that supports sustained rule adoption, diverse amendments, and balanced consensus. Smaller models tend to be rule-inert, while larger models often converge on restrictive voting patterns. Heterogeneous mixed-size groups collapse into veto-driven gridlock. These non-monotonic patterns persist under temperature perturbations and shifts from unanimity to majority voting, with the sweet spot showing linearly decodable vote-predictive information at intermediate-to-mid layers.

Key takeaway

For AI Scientists designing multi-agent LLM systems, recognize that institutional adaptability does not simply scale with model size. You should focus on identifying and utilizing mid-scale models, such as Qwen3.5 4B or Gemma3 12B, for tasks requiring complex collective adaptation and diverse rule dynamics. Be wary of deploying the largest models or heterogeneous groups, as they tend towards gridlock and restrictive behaviors, undermining the system's capacity for self-governance.

Key insights

Collective adaptation in LLM societies is non-monotonic, with mid-scale models exhibiting optimal rule dynamics.

Principles

• Optimal collective adaptation is scale-dependent and non-monotonic.
• Heterogeneous LLM societies tend to collapse into gridlock.
• Decodability of vote-predictive signals is necessary but not sufficient for adaptive play.

Method

The Nomic game serves as a controlled testbed for studying collective adaptation by allowing LLM agents to propose and vote on rule changes, coupling behavioral measurement with hidden-state probing.

In practice

• Avoid deploying largest LLMs for self-governing multi-agent systems.
• Prioritize mid-scale LLMs for emergent institutional dynamics.
• Ensure agent homogeneity for stable collective adaptation.

Topics

• Self-Amending LLM Societies
• Collective Adaptation
• Nomic Game
• Non-Monotonic Scaling
• Vote-Prediction Probing

Code references

• KazuyaHoribe/nomic-coevolution

Best for: AI Scientist, Research Scientist

Related on AIssential

• Scaling Behavior of Single LLM-Driven Multi-Agent Systems — LLM-driven multi-agent system performance shows diminishing returns due to coordination overhead, not monotonic scaling.
• RECAP: Regression Evaluation for Continual Adaptation of Prompts — Existing prompt adaptation methods are structurally inadequate for proactive, real-time constraint evolution in agentic systems.
• When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method — LLM-generated social networks are significantly influenced by prompt design, culture, language, and model scale.
• AI 101: Beyond RL: The New Fine-Tuning Stack for LLMs — Modern LLM fine-tuning is evolving into a modular stack, moving beyond monolithic RL to integrate diverse, efficient methods.
• Beyond Static Responses: Multi-Agent LLM Systems as a New Paradigm for Social Science Research — A six-tier framework maps LLM agent evolution from static tools to complex adaptive systems for social science research.
• Guiding LLM-Based Human Mobility Simulation with Mobility Measures from Shared Data — Guiding LLM-based mobility simulations with population-level measures improves collective behavior realism and efficiency.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.