acceptodds
Under review as a conference paper at ICLR 2027

Physically Grounded, Experience-Guided LLM Agents for Decentralized Grid Control

Abstract

Power grids are safety-critical infrastructure that must maintain reliable operation under changing demand and increasing shares of variable renewable generation. Topology control can relieve transmission congestion by reconfiguring substation busbars, but decentralized operation requires coordinating interventions across regions with only partial network information. Existing approaches lack a mechanism that combines assessment of the current physical state, coordination of interdependent regional actions, and evidence from past interventions within a single decentralized decision process. We propose GridRefine, a physically grounded, experience-guided framework for decentralized topology control using large language models as regional agents. Each agent uses a graph representation of its local network, physical knowledge, and a memory of past intervention outcomes from offline simulations to select and explain actions. Through counterfactual reasoning, agents consider how differences between past and current grid states may change an intervention’s effects, then refine their assessment of promising candidates using feedback from a local power-flow tool. Because these interventions can redistribute power flows across regional boundaries, neighbouring agents coordinate their proposals before execution. We evaluate GridRefine on two power grids of different sizes and include planned line disconnections for maintenance on the larger grid. With two of the three evaluated LLM backbones, GridRefine improves mean grid survival over MARL baseline across all three settings. Code will be released upon publication.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Large Language Model Agents, Power Grid Topology Control, Decentralized Multi-Agent Control, Counterfactual Reasoning, Physics-Grounded AI

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_GNfeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses the problem of multi-agent coordination of power grids under partial observability and cross regional coupling. It proposes GRIDREFINE, a decentralized topology-control framework for power grids in which each regional agent is powered by an LLM. In particular, GRIDREFINE's regional agents use historical outcomes and local physical calculations to select topology actions, coordinating when neighbouring proposals interact.

The framework is evaluated on Grid2OP 14-bus grid and a 36-bus grid (with and without planned maintenance), against Do-Nothing, MAPPO and a non-LLM retrieval controller. Three LLM backbones are tested: Gemma (4.5B), Qwen (27B) and DeepSeek (284B MoE). The results are compared to the MAPPO baseline (Yu et al., 2022) and demonstrate that QWEN and DeepSeek overcome MAPPO by 28–42 pp in mean survival.

The study report some ablations and detailed appendix to support their claims.

Critical strengths
  • Well motivated mechanisms through examples; for example in Fig. 2, where a historical failure overrides favourable DC estimates, is convincing.
  • Extensive appendices with sufficient reproducibility materials and transparent discussions of limitations and failures (of Gemma).
  • Insightful ablations and negative results, with sufficient discussions.
Critical weaknesses

The system is carefully engineered and thoroughly documented. The central empirical claims, however, rest on a weak, apparently unconverged baseline, missing natural baselines, and ablations with one seed and 20 episodes. The coordination mechanism is not convincingly shown to be the source of the gain.

  • The choice of MAPPO as sole baseline, and unfair comparison of MAPO. Figure 4 in appendix clearly shows MAPPO is not converged and can achieve significantly better results. Currently, MAPPO is below Do-Nothing on both Bus36 settings (44.7 vs 67.7; 27.6 vs 44.0), which is absurd, and confirm it was not properly/sufficiently trained.

  • The results are not significant:

-- Relative to Do Nothing, Qwen on Bus36 gains about 10.7 pp, totally within the standard deviation margins Qwen (61.4 ± 12.4) and Do-Nothing (44.0 ± 16.9).

-- Table 2 and Fig. 3 use 20 validation episodes and one seed (seed 0), no confidence intervals, and cumulative (not leave-one-out) ablations.

-- other qualitative claims.

  • Contradicting findings between main and appendix: Appendix D.3 reports that, over 7,788 coordinated events, there were 75 revisions, zero shared-line objections, zero fallbacks, and 1.01 average review cycles. Which means there is no benefit to the coordination; while the main ablation in Table 2 attributes +14.4 pp (Bus14) to coordination.

  • No demonstration of the generalization capabilities: All memory is grid-specific and built from the same data distribution (training chronics of the same grid). The test-time benefit of retrieval from 700/400 chronics may largely reflect closeness of test states to stored states on small benchmarks.

Questions for authors
  • Provide more up to date baselines: Including best single agent approaches (L2RPN winning agents) or more recent approaches like PowRL (https://arxiv.org/html/2212.02397v2), or decentralized coordination in power systems. Distributed/decentralized optimization for multi-area TSOs (ADMM, auxiliary problem principle, DCOP; see surveys such as Kargarian et al., 2016) is the established way to coordinate across control areas under partial information. The paper’s claim that “existing approaches lack a mechanism” for coordination ignores this literature. At least a discussion, and ideally a rule-based negotiation baseline, is needed.

  • Please validate the statistical significance of the claims, especially the qualitative ones (“coordination also improves survival with planned line disconnections”, “mining improves in all three settings”)

  • Explain the contradiction about the benefits of coordination.

  • Provide actual generalization results; for example: transfer to unseen topologies or grids; distribution shift (e.g., different seasonal profile, renewable penetration, enabled opponent/outages); robustness to memory size.

  • What is the test-time survival of the greedy AC-lookahead controller (with and without the one-region-per-step rule) on the same splits and catalogs? How many AC simulator calls does it use per decision compared with GRIDREFINE?

Reviewer_LHKbOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an LLM-based multi-agent framework, GridRefine that controls the electric grid in decentralized communication across multiple agentic regions. Each agent uses the network topology, physical knowledge, and historical information (memory) of its grid region. The framework is implemented by using MARL2Grid-TR. The numerical experiments were performed on small grid networks and compared with MAPPO. GridRefine shows substantial gains with respect to the survival metric.

Critical strengths
  • The topic is interesting and timely.
  • The paper reads well and presents an interesting approach.
  • The proposed framework has been designed and implemented on substantial domain knowledge with details.
  • The target use case (or situation) in grid operation might be of practical interest as well.
Critical weaknesses
  • The main contribution must be implementation and numerical experiments, lacking new technical development. Any new technical contributions should be clearly described. The paper currently reads very descriptive and verbatim.
  • The work needs strong justification for why each operating region would use an LLM-based agent. There are many existing work in distributed/decentralized grid controls. Why are the existing control/optimization not sufficient? What additional information should be communicated through LLM agents? What are the roles of using LLMs other than simple automation?
  • The work does not address the communication aspect, including topology and cost. The experiments consider up to four regions, which I assume that each agent is connected to everyone else. What if we have more agents with sparse communication topology? The grid control should be quickly responded, whereas the proposed approach may communicate large tokens per communication.
Questions for authors

Please see the questions in the weaknesses.

Reviewer_RapuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses decentralized power-grid topology control, where regional controllers have only partial observations and their interventions on shared lines can interfere. It proposes GRIDREFINE, in which each regional LLM agent chooses among candidate actions using retrieved past simulation outcomes, a check of whether those cases still apply to the current state, and a local DC power-flow estimate, with neighbouring agents reviewing each other’s proposals before execution. It is evaluated on 14-bus and 36-bus grids, and the two larger LLM backbones improve survival over MAPPO in all settings, while the smallest does not.

Critical strengths
  • The paper addresses a relevant, safety-critical problem with a sensible approach that is physically grounded and experience-guided.
  • The method improves survival over the baselines in every setting with the two larger LLM backbones.
Critical weaknesses
  • The ablation of the applicability assessment is missing. The paper presents assessing each retrieved case from differences between past and current states as the basis of its counterfactual reasoning (Section 4.1) but does not test it. The natural experiment would be the full method with the CASE TRANSFER step removed from the prompt.
  • The main baseline is under-trained. MAPPO, the only learned baseline, is still improving at 60M transitions (Figure 4) and is below Do-Nothing on both Bus36 settings (Table 1). The 60M budget is not justified.
  • The related work on RL for grid control covers benchmarks and multi-agent studies but not single-agent RL methods, such as [1,2,3].

[1] Lan et al., AI-based autonomous line flow control via topology adjustment for maximizing time-series ATCs, IEEE PESGM, 2020.

[2] Subramanian et al., Exploring grid topology reconfiguration using a simple deep reinforcement learning approach, IEEE Madrid PowerTech, 2021.

[3] Yoon et al., Winning the L2RPN Challenge: Power Grid Management via Semi-Markov Afterstate Actor-Critic, ICLR, 2021.

Questions for authors
  • Among retrieved cases the LLM labels strong, partial, or weak, how often does applying the same action at the current state in the simulator reproduce the recorded outcome?
  • Is the method expected to remain usable as the grid grows? Can it be applied to the 118-bus grid in MARL2Grid-TR?
Reviewer_rsfqOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an LLM framework, GridRefine, for topology control in a network shared/controlled by different TSOs, which do not have full observability of its neighbors' systems, and need to coordinate their operations. The proposed GridRefine framework is effectively an attempt to emulate the existing coordination process between TSOs to validate and agree on switching decisions in one TSO (or along the seam) that may impact neighboring TSOs. The coordination is simulated using separate regional LLM agents. The framework relies on DC power flow calculations to ensure physically-grounded results. The authors tested the framework on toy-size test systems: a 14-bus and a 36-bus test system.

Critical strengths

The problem tackled, topology control in shared power networks, is relevant and timely. The proposed LLM method is grounded in physical models.

Critical weaknesses

The paper does not read very well. The title is not descriptive of the topic actually discussed in the paper (e.g., grid control and topology control are not synonyms, the former is much more general) The authors, or the AI tools they used to write the paper, ignored the vast literature on both topology control and on inter-regional coordination over the last few decades. There is no justification for the proposed approach using LLM agents to tackle the topology optimization problem. Are the existing coordination approaches deficient, which requires this approach? Also, even in single-operator networks, why is LLM a good approach compared to physics-based optimization or similar heuristic approaches? Note that neither the model proposed (DC flows without contingency constraints) nor the systems used for the simulations (two to four orders of magnitude smaller than real systems) capture the complexity of either the topology optimization problem solved or the coordination challenges.

Questions for authors

See text in Critical Weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.