LLM-Mediated Demand Response Coordination in Smart Microgrids

AI summary

Overview

Research area: Multi-agent systems, large language model (LLM) agents, evolutionary game theory, and smart grid demand response.

Technical level: Advanced. The paper assumes familiarity with Prisoner's Dilemma payoff structures, Fermi-function imitation dynamics, scale-free network topology (Barabási–Albert), RLHF alignment behaviour, and multi-agent simulation design.

Scope: A simulation study testing whether an LLM-based "Influence Compiler" that issues structured demand-response directives can raise voluntary electricity-curtailment cooperation among heterogeneous prosumer agents more effectively than unstructured LLM messaging or no intervention at all.

What This Paper Is About

In a smart microgrid, households and businesses that both produce and consume electricity (prosumers) face a collective-action problem: everyone benefits if all cut demand during peak load, but each individual gains by free-riding on others' cuts. The authors model this as a repeated Prisoner's Dilemma played on a social network and ask whether an LLM-powered coordinator can nudge prosumers toward cooperation. The obstacle they identify is that RLHF-aligned LLMs are so cooperative by default that they flatten every experimental condition into identical behaviour — so the paper builds a hybrid architecture where game theory supplies the strategic tension and the LLM only evaluates the persuasive message.

Key Contributions

1. A hybrid decision architecture (v3) that separates a game-theoretic base cooperation probability (from payoff history, neighbourhood imitation, exploitation memory, and temptation decay) from an LLM narrative evaluation layer whose output is bounded to [−0.30, +0.30] and attenuated by agent resistance. This prevents RLHF cooperation bias from collapsing experimental conditions.

2. An LLM Influence Compiler that emits structured policy schemas — target, intensity ∈ [0,1], timing ∈ {burst, periodic}, and theme ∈ {moral, economic, identity, hybrid} — rendered into natural language via a Solver–Critic pipeline, rather than generating free-form persuasive text.

3. A six-archetype prosumer population (pragmatist, idealist, skeptic, conformist, strategist, opportunist) with distinct base biases and resistance levels, producing realistic behavioural heterogeneity with baseline cooperation near 50%.

4. A four-part experimental evaluation covering compilation effect, grounded vs. idealized substrates, network targeting strategies, and a resistance sweep, plus agent-level decomposition of persuasion and backlash.

Main Findings

• Compiled directives beat unstructured messaging and the baseline. Compiled influence reached a final cooperation rate of 0.333 (time-averaged 0.305), versus 0.270 (0.275) for unstructured messaging and 0.280 (0.282) for no influence. The compilation advantage Δ_comp was +0.063, with a lift above baseline of +0.053. Unstructured messaging performed below the no-intervention level, i.e., a slight backfire.

• The advantage survives realistic agent behaviour. Idealized logistic responders reached 0.850 cooperation under compiled influence and 0.767 under unstructured (Δ = +0.083), while grounded hybrid agents reached 0.333 and 0.270 (Δ = +0.063). The absolute gap between substrates was −0.517, but the relative ordering compiled > unstructured > no-influence held in both cases.

• Hub and bridge targeting dominate. Under compiled influence with grounded agents, HUBS and BRIDGES both produced 0.333 final cooperation and 0.305 average; PERIPHERY produced 0.307 final and 0.313 average; RANDOM produced 0.270 final and 0.275 average. Periphery's higher average but lower final rate suggests temporary local clustering among low-degree prosumers that did not propagate.

• The advantage never vanishes under resistance. Sweeping R ∈ {0.1, 0.3, 0.5, 0.7}, compiled cooperation stayed within a range of only 0.007 (0.330, 0.333, 0.333, 0.327), and Δ_comp remained strictly positive at every level (+0.067, +0.063, +0.070, +0.057). The compiled condition at R = 0.7 (0.327) still exceeded the unstructured condition at R = 0.1 (0.263).

• Personality archetypes differentiate meaningfully. Final cooperation propensity was highest for idealists (0.447), then conformists (0.383) and pragmatists (0.379), and lowest for opportunists (0.224), skeptics (0.246), and strategists (0.272). This differentiation is presented as evidence that RLHF bias is not dominating the behaviour.

• Persuasion collapses with resistance. Agents with r < 0.20 (mainly idealists and conformists) were persuaded in roughly 85–100% of targeted interactions, while agents with r > 0.40 (skeptics, opportunists) were persuaded in fewer than 5%. Skeptics registered zero persuasion events because they were never targeted, given their high resistance penalty in the compiler's targeting score.

• High-degree agents trend more cooperative. The degree-versus-strategy plot shows hub prosumers tending toward modestly higher cooperation, consistent with the conformist imitation term receiving more cooperative neighbourhood signals.

• Persuasion frequency is not the mechanism. Persuasion rates fluctuated around 0.50 for both influence conditions, so the cooperation gap is attributed to the quality of the LLM shift rather than how often it was applied.

Methodology in Plain English

The authors built a simulated energy community of 30 prosumer agents on a scale-free network generated by Barabási–Albert preferential attachment with m = 3, and ran it for 50 discrete time steps. Each step, every agent chooses to curtail demand (cooperate) or consume normally (defect), with payoffs set to the classic Prisoner's Dilemma values R = 3, T = 5, S = 0, P = 1.

Each agent's cooperation probability has two parts. The first is a logistic function combining an archetype bias, an exploitation count, the neighbourhood's mean cooperation rate, the payoff advantage of defection in the last round, and a temptation-decay term, using fixed weights {0.5, 1.2, 0.8, 0.6, 0.3}. The second is triggered only when the agent is targeted by the central Influence Compiler.

The compiler, which models a demand-response aggregator or distribution system operator, sees a noisy snapshot of the population (Gaussian noise N(0, 0.10)) every 5 steps and sends a directive to ⌊0.2N⌋ = 6 agents. It picks a structured policy (target, intensity, timing, theme) and renders it into language through a Solver–Critic loop. Instead of asking the LLM to decide cooperate-or-defect — which the authors say produces near-universal cooperation — the agent is prompted under a NARRATIVE_EVAL_SYSTEM instruction to output only how much its willingness to curtail changes, as a scalar in [−0.30, +0.30]. That shift is multiplied by (1 − r_o) for the agent's resistance and added to the base probability, then the final probability is clipped to [0.02, 0.98] and sampled as a Bernoulli action.

Non-targeted agents make no LLM call; they may receive an attenuated word-of-mouth shift drawn from N(0, 0.05) if a cooperating neighbour transmitted the directive. Targeting strategies compared are HUBS (top-k by degree), BRIDGES (top-k by betweenness), PERIPHERY (bottom-k by degree), and RANDOM. All LLM calls used Llama-3.3-70B-Instruct via the Nebius AI Studio API with SHA1-based caching for reproducibility.

Results are averaged over the final ten time steps (t ∈ [40, 49]) to reduce stochastic noise. The headline metric is the population cooperation rate; secondary metrics are final strategy, persuasion rate (share of deployments with positive LLM shift), and backlash rate (share with negative shift).

Why This Matters

Impact on research: The paper argues that RLHF-aligned LLMs are unsuitable as direct decision-makers in controlled cooperation experiments because their prosocial bias erases treatment effects. Its hybrid design offers a reusable template for any multi-agent study that wants LLM reasoning without LLM bias dominating the results, and it links the LLM multi-agent literature to the established game-theoretic demand-response literature.

Real-world applications:

• Utilities and demand-response aggregators issuing curtailment requests during peak load events.
• Urban energy communities and microgrid operators coordinating voluntary flexibility without coercive contracts.
• Network-aware outreach programmes that prioritize high-centrality participants (hubs and bridges) over broad or peripheral campaigns.
• Governance design for influence systems, where Constitutional AI-style constraints could filter fear-based framing or biased subgroup targeting.

Industry relevance: The paper's three stated design principles — use structured directive schemas rather than generic messaging, prioritize network hubs and bridges for dissemination, and apply constitutional filters to prevent manipulative directives — are directly actionable for smart grid communication platforms. The finding that unstructured messaging performs slightly worse than no messaging is a caution for operators assuming any engagement is beneficial.

Future Directions

• Human-subject validation. The authors identify validation with real prosumers in real energy community settings as the primary direction for future work, since current agents are LLM proxies for humans.

• Scaling up. Larger populations and heterogeneous grid topologies beyond the single scale-free network with N = 30 are listed as open extensions.

• Governance mechanisms. Filtering directives through constitutional or normative constraints is proposed to prevent manipulative or inequitable messaging, but is not implemented here.

• Open questions raised by the results: why periphery targeting showed a higher average but lower final cooperation rate than hub targeting, and whether the strictly positive compilation advantage persists outside the tested resistance range R ∈ [0.1, 0.7].

Target Audience

Researchers working on LLM multi-agent systems, evolutionary game theory, and cooperative AI will get the most from the methodological argument about RLHF bias and the hybrid architecture. Smart grid and demand-response researchers, energy community operators, and smart city planners will find the targeting and message-structure findings directly applicable. Practitioners building agent-based simulations of social coordination will benefit from the concrete simulation parameters, archetype definitions, and evaluation metrics. The paper's stated limitations (N = 30, single network realisation per condition, LLM agents as human proxies) mean readers seeking validated field results should treat it as a design and simulation contribution rather than an empirical grid study.

Authors’ abstract

Effective demand response in smart microgrids requires prosumers to cooperate voluntarily under strategic self-interest, a coordination problem structurally equivalent to a repeated Prisoner's Dilemma on a social network. This paper presents a multi-agent simulation in which a Large Language Model (LLM) Influence Compiler issues structured demand-response directives to a population of heterogeneous prosumer agents, each governed by a hybrid decision architecture combining game-theoretic base probability (derived from payoff history, neighbour imitation, and exploitation memory) with LLM narrative evaluation of incoming coordination signals. The hybrid architecture resolves a key methodological challenge: LLMs aligned via Reinforcement Learning from Human Feedback (RLHF) exhibit strong cooperation bias when used as direct decision-makers, producing flat dynamics regardless of grid conditions. By separating strategic reasoning from grounded narrative evaluation, the model generates realistic prosumer behaviour across six personality archetypes, with baseline cooperation near 50% and clear differentiation under influence. Compiled structured directives achieve 33.3% demand-curtailment cooperation versus 27.0% for unstructured messaging and 28.0% for a no-intervention baseline ($Δ_\mathrm{comp} = +0.063$), with the advantage preserved across both grounded and idealized agent substrates ($Δ= +0.083$) and across all resistance levels ($R = 0.1$ to $0.7$). Hub-targeted dissemination via high-centrality network nodes outperforms peripheral or random targeting, confirming that grid topology provides mechanistic amplification independent of message content. These results suggest 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.