LLM Influence Compiler boosts smart microgrid cooperation by 33%

The paper tackles a classic coordination problem in smart microgrids: getting prosumers (energy producers/consumers) to voluntarily curtail demand during peak periods, even when it's against their individual short-term interest. The authors frame this as a repeated Prisoner's Dilemma on a social network and introduce an LLM-mediated system. Their key innovation is a hybrid agent architecture where a game-theoretic module handles strategic reasoning (based on payoff history, neighbour imitation, and exploitation memory) while an LLM evaluates incoming structured directives from a central 'Influence Compiler' separately. This avoids the well-known cooperation bias of RLHF-aligned LLMs when used as direct decision-makers. The simulation models six personality archetypes and shows that the system produces realistic baseline cooperation near 50% with clear differentiation under influence.

Results demonstrate that structured, compiled directives achieve 33.3% demand-curtailment cooperation, beating unstructured messaging (27.0%) and a no-intervention baseline (28.0%) — a relative improvement of 6.3 percentage points. The advantage holds across both idealized and more realistic 'grounded' agent substrates (Δ = +0.083), and across varying resistance levels. Importantly, the study finds that hub-targeted dissemination (directing messages to high-centrality nodes in the network) outperforms random or peripheral targeting, confirming that grid topology provides a mechanical amplification effect independent of message content. The paper concludes that structured LLM compilation, grounded agent reasoning, and network-aware targeting are complementary design principles for scalable, interpretable demand-response coordination in smart-city energy systems.