Source-linked AI summary
Grounded, Compute-Efficient LLM Policy Agents for Energy-Poverty Equity in Physically-Constrained Peer-to-Peer Energy Markets
Kunal Jadhav, Siddhesh More
TL;DR
Energy-poverty research in NLP has limited coverage and often relies on static systems or carbon-intensive cloud LLMs. EqGrid combines grounded household personas, an LLM policy layer, MARL energy trading, and grid-safe execution. It reports improved equity and cost outcomes, substantial retained benefit in smaller models, and zero grid violations through decoupled safety.
Problem
Energy poverty is underrepresented in NLP for social good, while existing work is largely static retrieval/QA or uses carbon-intensive cloud LLMs.
Method
EqGrid simulates a closed-loop energy market with grounded household personas, a low-frequency open-weight LLM policy agent, MARL traders, and a validate-and-project grid gate.
Results
0.305 versus 0.351: the LLM policy lowers the Gini of energy burden, while a 3B-active model retains 95% of the equity gain at approximately 9× lower energy and the safety gate yields zero violations versus 55.
Takeaways & Limitations
Energy-poverty equity and the AI system’s own energy cost can be evaluated together, with much of the benefit retained by smaller open models.
Takeaways & Limitations
Persona grounding uses Hungarian EU-SILC marginals rather than restricted microdata, and per-decision energy figures are estimates rather than physical-hardware measurements.
Abstract
from arXiv · showhide
Energy poverty is nearly absent from NLP-for-social-good, and the little existing work is either static retrieval/QA or relies on carbon-intensive cloud LLMs, a self-defeating "computational irony" for a humanitarian setting. We present EqGrid, a closed-loop simulation in which a low-frequency, open-weight LLM policy agent sets price and carbon bounds and targeted subsidies over a community of empirically-grounded household personas, while high-frequency multi-agent RL traders clear a continuous double auction constrained by a physical distribution grid (IEEE-33-bus with Dynamic Operating Envelopes). Our contribution is threefold and directly addresses how to measure the social impact of AI: (i) grounded personas (region-matched socio-demographics) whose load curves are checked for shape and level realism against real smart-meter data; (ii) formal energy-poverty equity metrics (Energy Burden, Gini of EB, LIHC) showing the intervention reduces burden inequality without raising net grid cost; and (iii) a compute-efficiency frontier that measures how much equity performance survives compressing the policy agent from a 235B teacher down to a sub-1B model deployable on a laptop, in estimated energy/carbon per decision. A decoupled-safety design (the LLM sets bounds; a validate-and-project grid gate executes) yields zero grid-constraint violations versus 55 under direct LLM control. On energy-poverty equity, the LLM policy lowers the Gini of energy burden to 0.305 (from 0.351) and mean burden by 28% while cutting cost (outperforming a tuned rule baseline), and a 3B-active model retains 95% of the benefit at roughly 9x lower inference energy than the teacher, with even a 0.8B on-device model retaining 92% at roughly 24x lower energy. We will release code and configs.
1 Introduction
EqGrid integrates grounded household personas, a compute-efficient LLM policy layer, MARL market trading, and a physically constrained grid to measure energy-poverty equity alongside AI cost and safety.
- The paper positions EqGrid as an integrated framework for energy-poverty mitigation and measurement of AI’s social impact.The authors state that the contribution lies in the integrated system and measurement framework rather than an individual algorithmic component.
- EqGrid combines grounded household personas, an open-weight LLM policy agent, MARL double-auction traders, and an IEEE-33-bus grid with Dynamic Operating Envelopes.The LLM sets price and carbon bounds and targeted subsidies while MARL traders clear the market.
- 0.305 versus 0.351: the LLM policy lowers the Gini of energy burden while reducing cost and outperforming a transparent rule baseline.The measurement suite includes energy burden, Gini of energy burden, LIHC prevalence, and carbon-adjusted SROI.
- 92% of the equity gain survives compression to a 0.8B laptop-deployable model at approximately 24× less energy per decision.The compute frontier measures retained social benefit against the policy agent’s energy and carbon cost.
- 0 grid violations versus 55 under direct LLM control: a validate-and-project gate executes trades safely while the LLM sets bounds.The safety design separates language-model policy decisions from physical control.
- Code, configurations, and the measurement harness will be released upon publication.
2 Related Work
Related work spans energy-domain NLP, socioeconomic LLM agents, MARL P2P markets, and Green AI, but EqGrid connects these strands to grounded, physically constrained energy-poverty intervention.
- Existing energy-domain NLP is largely informational, while EqGrid simulates interventions on energy poverty.Surveys identify poverty and energy as among the least-addressed social-good domains in the ACL Anthology.
- LLM socioeconomic-agent work grounds populations in real data but has not addressed energy or physical grids in this setting.EqGrid adapts planner-over-population designs to a physically constrained energy market with energy-poverty equity and edge efficiency.
- MARL and LLM-guided P2P-market studies optimize generic fairness or cost while omitting grounded personas, edge efficiency, and physical network constraints.Prior systems also couple the language model to the fast trading loop.
- Inference energy varies by orders of magnitude across models, motivating measurement of AI footprint and small, open-weight, on-device backbones.The paper reports no prior energy-poverty system measuring the carbon cost of its own AI.
- Figure 1 depicts the closed loop in which persona grounding, MARL trading, economic bounds, and the physical grid are connected, with safety enforced by a validate-and-project gate.
3 The EqGrid Framework
EqGrid closes a grounded household simulation around a low-frequency policy agent, high-frequency MARL market, and deterministic physical-grid safety layer, while tracing equity retained against compute cost.
- Four components form the closed loop: grounded household personas, an LLM policy agent, MARL traders, and a physical distribution-grid safety layer.The LLM sets market bounds but never executes physical actions.
- Households are instantiated from Hungarian EU-SILC marginals, converted into activity profiles and hourly loads by an LLM plus a deterministic appliance-and-heating model.Resulting load distributions are checked for shape and level realism against real smart-meter data.
- Every 6-hour policy period, the agent observes community equity and tariff summaries and returns JSON price bounds, a subsidy weight, and carbon allowances.The policy targets lower energy-poverty inequality without raising net grid cost while keeping exporters viable.
- MARL household traders control home batteries and bid prices in hourly continuous double auctions bounded by the policy band.A compact MAPPO uses a shared-parameter actor and centralized critic; unmatched volume settles with the utility.
- 3.4 Physical grid layer and decoupled safety: The validate-and-project gate checks aggregated injections against voltage and thermal limits, projecting infeasible proposals by global curtailment.This design guarantees zero violations while curtailing as little as possible.
- 3.5 Compute-efficiency frontier: The compute frontier varies the policy backbone from a 235B teacher to a sub-1B laptop model while holding persona generation fixed.Each model rung is evaluated by equity outcome and energy/carbon per policy decision.
4 Measuring Social Impact: Metrics
The paper measures both energy-poverty outcomes and the policy agent’s computational footprint. Its metrics assess burden, inequality, vulnerable households, social return, and energy/carbon per decision.
- Energy Burden measures each household’s net annual energy cost relative to income.
- Gini of Energy Burden measures community burden inequality, with lower values preferred without raising net grid cost.
- LIHC prevalence counts below-median-income households whose energy share exceeds 10%, while the 90th-percentile burden tracks vulnerable households and hidden energy poverty.
- Carbon-adjusted SROI subtracts monetized LLM-inference carbon from monetized social value before dividing by initial investment.
- Energy/carbon per decision estimates inference footprint from generated tokens and active parameters, using a calibrated proxy rather than on-device power measurements.
5 Experiments
The experiments simulate a grounded Hungarian residential community in a physically structured energy market and compare policy, grounding, compute, and safety conditions. Evaluation uses five paired seeds and includes established market baselines and equity-table reporting.
- Setup: The setup models a Budapest community on an IEEE-33-bus feeder over 24 winter hours, with hourly auctions and 6-hourly policy periods.
- Setup: Personas use Hungarian EU-SILC marginals, and generated load curves are checked against a real smart-meter reference for shape and level realism.
- Protocol and statistics: Five paired seeds support mean ± standard deviation reporting, paired significance tests, 95% confidence intervals, and paired effect size dz.
- Conditions: The conditions vary policy against NONE and RULE, grounded versus synthetic personas, policy-model size, and DOE-gated versus direct LLM control.
- Baselines: Baselines include a no-policy MARL market, a transparent rule-based policy, and a large cloud model as a carbon-intensive reference.
- Results: Table 1 reports equity by policy as mean ± standard deviation over five seeds, where lower Gini, burden, LIHC, and bill are better.
6 Results
EqGrid’s results show that small, open models preserve most energy-poverty equity gains at far lower inference energy, while safety gating eliminates grid violations. The LLM policy improves equity and cost outcomes, though grounded-versus-synthetic persona effects remain inconclusive.
- Equity: 0.305 Gini of energy burden, versus 0.330 for the rule and 0.351 for no policy, was achieved alongside the lowest mean burden, low-income-high-cost prevalence, and daily cost.The LLM significantly reduced Gini against no policy and significantly reduced cost against both no policy and the tuned rule; its additional Gini reduction over the rule was not robustly significant at five seeds.
- Grounding ablation: Grounding improved subsidy-to-burden alignment numerically, but the difference was not statistically significant and absolute equity levels were not directly comparable across populations.Grounded personas reproduced realistic socioeconomic structure and load-curve shape and level, but decisive comparison requires restricted EU-SILC microdata and a multi-household benchmark.
- Compute-efficiency frontier: 95% equity retention at 0.029 Wh per decision made the 3B-active MoE model a compute-efficient alternative to the 235B teacher’s 0.27 Wh reference.The model ladder measures retained equity against energy per policy decision.
- Compute-efficiency frontier: 92% equity retention at 0.011 Wh per decision enabled the 0.8B laptop-deployable model to retain most of the teacher’s equity gain at approximately 24× lower energy.The fully open olmo3-7b also retained 92% at 0.078 Wh, while gemma4-e2b retained 85% at 0.026 Wh.
- Compute-efficiency frontier: The dense qwen36-27b model’s 31% retention reflected truncation-induced fallback, with 90% of decisions lacking a closeable JSON output.The model exhausted its generation budget on chain-of-thought, making the failure deployment-relevant rather than evidence of poor general reasoning.
- Safety: 0 grid violations occurred with the validate-and-project gate, compared with 55 under direct LLM control.The gate executes feasible actions while the LLM only sets bounds.
7 Limitations
The study’s limitations concern persona and load validation, inference-energy measurement, and subsidy funding. These constraints narrow how its grounding, compute, and equity claims should be interpreted.
- Persona and load validation: EU-SILC grounding uses published Hungarian marginal distributions rather than restricted microdata, reproducing marginal and correlational structure rather than a specific sample.The authors identify multi-household, region-matched distributional validation as an immediate next step.
- Persona and load validation: Load validation relies on a single French household dataset from 2006–2010, so it is only a coarse check of diurnal shape and consumption level.It is not a region- or period-matched validation of Hungarian aggregate loads.
- Compute measurement: Per-decision energy figures are active-parameter-based estimates calibrated to literature rather than measurements from physical hardware.The authors state that on-device profiling with hardware power counters or CodeCarbon remains future work.
- Subsidy design: The targeted subsidy is an unfunded transfer without a levy, budget cap, or revenue-neutrality constraint, so it redistributes burden without a modelled funding source.Across eight policy conditions, Gini of energy burden is largely determined by the emitted subsidy-weight scalar, with R2=0.985.
Ethical and Societal Implications
The paper defines positive impact through measurable energy-poverty equity gains that do not increase net cost or impose outsized environmental cost. It pairs this objective with subgroup checks, deterministic safety controls, local data handling, and deployment cautions.
- Impact definition: Positive impact means reducing energy-burden inequality without raising net cost or imposing an outsized environmental cost.The framework operationalizes this with Gini of energy burden, LIHC prevalence, the vulnerable tail, and carbon-adjusted SROInet.
- Fairness and disaggregated harm: Demographic-parity reporting and grounding ablations check whether support reaches genuinely high-burden households across income groups.The paper gives particular scrutiny to multigenerational and low-efficiency-dwelling households.
- Safety: The LLM sets only economic bounds, while a deterministic power-flow gate projects trades onto the safe operating envelope, yielding zero grid violations.No physical action is taken from an unverified LLM output.
- Privacy: Small open-weight on-device models keep household data local, avoid third-party cloud inference, and expose the policy agent only to aggregate community statistics.Individual consumption traces are not shown to the policy agent.
- Risks and deployment: Because results are simulated, the authors caution against deployment without field validation and participation from affected communities and utilities.Code release and an explainable rule baseline are presented as governance and transparency mitigations.
A Model Identifiers and Licenses
The reproducibility appendix identifies every model used on the compute-efficiency ladder and describes how model openness and active parameters support the energy estimates.
- Model identifiers and licenses: Table 3 provides each model’s public checkpoint, parameter count, and license; all six models are open-weight and publicly obtainable.OLMo-3 additionally releases its training data and code.
- Model identifiers and licenses: Models were served unmodified through an OpenAI-compatible API, and each model’s active parameter count drives its per-decision energy estimate.
B Literature Matrix
The literature matrix positions EqGrid against related work by identifying shared components and the capabilities those systems lack relative to the paper’s integration.
- Literature matrix: Table 4 compares EqGrid with 2021–2026 work across seven research threads, with each row sharing one or more components and the final column identifying missing integration elements.
- Literature matrix: EqGrid is presented as the first system jointly coupling grounded energy-poverty personas, a compute-efficient LLM policy, MARL trading, and a physical IEEE-33/DOE grid.Its evaluation combines formal energy-poverty equity metrics with the AI system’s own carbon cost.
- Literature matrix: Table 3 is cited in the appendix as documenting the open-weight models and licenses underlying the compute-efficiency ladder.