LLM-Mediated Demand Response Coordination in Smart Microgrids

Technical Summary: LLM-Mediated Demand Response Coordination in Smart Microgrids

Problem Statement
The paper addresses the challenge of coordinating voluntary demand response in smart microgrids, where "prosumers" (agents who both produce and consume energy) must decide whether to curtail demand during peak loads. This scenario is modeled as a repeated Prisoner's Dilemma on a social network: collective welfare is maximized through universal cooperation, but individual incentives favor defection (free-riding). While Large Language Models (LLMs) offer potential for generating interpretable, adaptive coordination directives, a critical methodological flaw exists: LLMs aligned via Reinforcement Learning from Human Feedback (RLHF) exhibit a strong prosocial bias. When used as direct decision-makers, they produce near-universal cooperation regardless of grid conditions or incentive structures, thereby collapsing experimental dynamics and obscuring the effects of different coordination strategies.

Methodology
To resolve the RLHF bias while leveraging LLM capabilities, the authors propose a hybrid decision architecture within a multi-agent simulation of a smart microgrid ($N=30$ agents on a scale-free network).

1. Hybrid Decision Architecture: The model decouples strategic reasoning from narrative evaluation.

   • Stage 1 (Game-Theoretic Base): An agent's baseline cooperation probability ($p_{base}$) is calculated using evolutionary game theory principles. It aggregates payoff history, exploitation memory (past rounds where the agent cooperated while neighbors defected), neighbor imitation (conformism), and a temporal temptation decay term. This ensures a baseline cooperation rate near 50%, creating genuine strategic tension.
   • Stage 2 (LLM Narrative Evaluation): Instead of asking the LLM to make a binary decision, the system prompts the LLM to act as an "Influence Compiler" that evaluates incoming directives. The LLM outputs a scalar shift ($\delta_{LLM} \in [-0.30, +0.30]$) representing how much the agent's willingness to curtail changes based on the directive's narrative, the agent's persona, and history.
   • Stage 3 (Resistance Attenuation): The final cooperation probability is a clipped sum of the base probability and the LLM shift, attenuated by an agent-specific resistance factor ($r$). This prevents the LLM from overriding the strategic base entirely.

2. Agent Archetypes: The simulation includes six heterogeneous prosumer archetypes (Pragmatist, Idealist, Skeptic, Conformist, Strategist, Opportunist), each with distinct base biases and resistance levels.

3. Influence Compiler: A central compiler observes a noisy snapshot of the grid state and issues structured directives (target, intensity, timing, theme) to 20% of agents every five time steps. The compiler uses a Solver–Critic pipeline to generate structured policies, contrasting with an unstructured baseline using free-form narrative generation.

4. Targeting Strategies: The study evaluates dissemination via high-centrality hubs, bridges, peripheral nodes, and random selection.

Key Contributions

• Methodological Innovation: The paper introduces a hybrid architecture that separates game-theoretic reasoning from LLM narrative evaluation. This prevents the "cooperation collapse" caused by RLHF bias, allowing for the measurement of genuine treatment effects in coordination experiments.
• Structured vs. Unstructured Influence: It demonstrates that structured policy compilation significantly outperforms unconstrained narrative generation in driving demand curtailment.
• Network Topology Effects: The study confirms that targeting high-centrality nodes (hubs and bridges) provides mechanistic amplification of coordination signals, independent of message content.
• Robustness Analysis: The framework is tested across varying agent resistance levels and between "grounded" (hybrid) and "idealized" (logistic) agent substrates.

Experimental Results

• Cooperation Rates: The compiled structured directives achieved a final cooperation rate of 33.3%, compared to 27.0% for unstructured messaging and 28.0% for a no-intervention baseline. This yields a compilation advantage ($\Delta_{comp}$) of +0.063. Notably, unstructured messaging performed slightly worse than the no-intervention baseline, suggesting a "backfire effect."
• Substrate Robustness: The advantage of structured compilation persisted in both grounded (hybrid) and idealized agent models. While idealized agents reached higher absolute cooperation (0.850 vs. 0.333), the relative ordering (Compiled > Unstructured > No-Influence) remained consistent.
• Resistance Sweep: The compilation advantage remained strictly positive across all tested resistance levels ($R \in [0.1, 0.7]$), ranging from +0.057 to +0.070. The advantage did not vanish even at high resistance, indicating graceful degradation rather than threshold collapse.
• Targeting: Hub-targeted dissemination outperformed random and peripheral targeting. The periphery strategy showed only temporary local clustering that failed to propagate through the network.
• Agent Heterogeneity: The hybrid architecture successfully preserved personality differentiation. Idealists showed the highest persuasion rates (~85–100%), while high-resistance agents (Skeptics, Opportunists) showed rates below 5%. This contrasts with pure LLM models where all agents would converge to high cooperation.

Significance and Claims
The paper claims that structured LLM compilation, grounded agent reasoning, and network-aware targeting are complementary design principles for scalable demand-response coordination. The primary significance lies in solving the methodological challenge of using RLHF-aligned LLMs in cooperation research; by isolating the LLM's role to narrative evaluation rather than decision-making, the model generates realistic prosumer behaviors that reflect strategic self-interest.

The authors conclude that for next-generation smart city energy systems:

1. Communication systems should employ structured directive schemas rather than generic messaging.
2. Dissemination strategies must prioritize network hubs and bridges.
3. Constitutional governance constraints can be applied to filter manipulative directives with minimal cost to cooperation.

The paper remains modest regarding its scope, acknowledging limitations such as the small population size ($N=30$), single network realization, and the use of LLM agents as proxies for humans. It identifies validation with human subjects in real energy communities and scaling to larger, heterogeneous topologies as necessary future work.