From LLM narratives to parameterized cooperation policies in multi-agent systems

Area of Science:

• Artificial Intelligence
• Computational Social Science
• Agent-Based Modeling

Background:

• Large language models (LLMs) can generate persuasive narratives to shift agent behavior in multi-agent systems.
• However, raw, unstructured text lacks formal guarantees for effectiveness, interpretability, or controllability.

Purpose of the Study:

• Introduce the LLM Influence Compiler (LIC), a solver-critic pipeline.
• Compile natural-language cooperation directives into structured, parameterized influence policies.
• Provide formal guarantees for LLM-driven influence in artificial societies.

Main Methods:

• The LIC compiles directives into policies with schema: network targeting, narrative theme, intensity, deployment timing, and compiler confidence.
• Policies are diffused via an exposure model with fatigue decay, susceptibility heterogeneity, and backlash.
• Evaluated across 200+ simulations, 4 topologies, 4 network sizes, 5 LLM backbones, and 2 non-stationary perturbations.

Main Results:

• Compiled policies increased mean cooperation rate to 0.826 ± 0.010, an 8.6% relative improvement over baselines.
• Controlled decomposition attributed gains to network targeting (+4.5 pp), dose calibration (+1.9 pp), and intervention floor (+1.5 pp).
• LLM-mediated re-deployment significantly outperformed frozen rules under non-stationarity (p < 0.001).

Conclusions:

• The compiler abstraction converts opaque LLM processes into decomposable, auditable policy objects.
• This enables independent attribution, comparison, and governance of LLM-mediated influence.
• Lays groundwork for constitutional oversight of LLM influence in artificial societies.