Accepted · NLP4PI @ EMNLP 2026

Grounded, Compute-Efficient LLM Policy Agents for Energy-Poverty Equity in Physically-Constrained Peer-to-Peer Energy Markets

NLP4PI (NLP for Positive Impact) Workshop @ EMNLP 2026, 2026

Summary

EqGrid is 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 reinforcement learning traders clear an hourly double auction on a physically constrained IEEE 33-bus grid. A deterministic validate-and-project safety gate enforces the physics: it reduces voltage violations from 55 under direct LLM control to 0. The LLM policy lowers inequality and cost together (Gini of energy burden 0.351 to 0.305, bills down 27.7 euros per day). A compute ladder over six models shows that bigger is not better for low-frequency control: a 0.8B-parameter model retains 92% of the equity benefit at roughly 24x less energy per decision.