acceptodds
Under review as a conference paper at ICLR 2027

Solver Success Is Not Enough: Verification-Grounded LLM Agents for Distribution-Grid Model Reconstruction

Abstract

Updating a distribution-grid optimization model from a natural-language operational request is not a code-generation task alone. A single request may jointly change device states, parameters, constraints, objectives, and priorities, creating dependencies across interpretation, formulation, implementation, and execution. A candidate may execute and solve successfully while misrepresenting the request or producing a solution that violates nonlinear AC power-flow constraints. Existing evaluations of LLM agents do not jointly test such coupled modifications against both requirement fidelity and physical feasibility. We introduce a benchmark with 36 request types spanning parameter edits, formulation changes, and coupled state-and-formulation changes. To evaluate the same reconstruction tasks across different network structures, we instantiate all 36 request types under four operating conditions on three distribution networks, yielding 432 cross-network instances, each paired with executable requirement and physical checks. We also propose GridRecon, a verification-grounded multi-agent framework in which Planning, Formulation, Coding, and Review Agents exchange structured requirement specifications and mathematical change plans. Deterministic tools check requirement implementation, program execution, topology, AC power flow, and SOC relaxation quality, while a controller uses the resulting evidence to accept verified candidates, trigger bounded, targeted self-refinement for actionable failures, or reject unresolved cases. As a result, our GridRecon outperforms the strongest adapted baseline for each of four LLM backbones on IEEE 69 by 7.6 to 34.0 percentage points and passes 329 of 432 instances in the common cross-network evaluation on IEEE 33, IEEE 69, and case136ma.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, Automated Optimization Modeling, Power Distribution Systems

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ZmU9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a multi-agent (specifically, 4-agent) system to solve the update requests to power grids.

It proposes new benchmarks and the proposed method shows consistent improvement compared to other LLM-based methods.

Critical strengths
  1. The overall writing is good and clear
  2. The model design is reasonable
  3. It proposes new task benchmarks on existing grid networks.
Critical weaknesses
  1. Looks like there exists related benchmark: “ProOPF: Benchmarking and Improving LLMs for Professional-Grade Power Systems Optimization Modeling”; maybe more experiments shall be conducted on the existing benchmarks. Right now this paper proposes new benchmark and show performance gain only on the proposed benchmark is not fully convincing.
  2. All the baselines are LLM-based. I m not super familiar with related works in this domain but there shall include non-LLM-based methods. Comparison with them would be important.
  3. Overall in my view, the proposed method is a domain-specific way to combo the existing LLMs with appropriate prompt engineering works. That level of novelty is at most on the borderline for a top-tier ML conference.
Questions for authors

Please check the weaknesses mentioned above.

Reviewer_X2KTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

The paper introduces GridRecon, a multi-agent framework for modifying distribution-grid optimization models from natural-language requests. It organizes planning, formulation, coding, and review into a structured workflow with tools and bounded feedback-driven revisions. Candidate solutions are checked for requirement fidelity, numerical solvability, and physical feasibility. The paper also introduces a benchmark covering parameter edits, formulation changes, and coupled modifications across three distribution networks. Experiments with four LLM backbones show higher end-to-end success rates than the compared frameworks.

Critical strengths
  1. The paper addresses a real problem. A program can execute and solve successfully while failing to implement the user’s request or satisfy physical constraints. Evaluating these aspects separately is practically valuable.

  2. The paper is easy to follow. The motivation and workflow are clearly explained, making the approach accessible to readers.

Critical weaknesses
  1. GridRecon emphasizes its multi-agent design, but it essentially uses a fixed workflow with role-specific prompts for the same LLM. This division of roles and revision process is not new.

  2. The paper also does not sufficiently justify splitting the system into four separate agents. A useful alternative would be a single agent harness that exposes the relevant operations and modeling guidance as tools and skills, allowing the model to decide when to use them. The appendix describes the OpenCode setup, but explicitly states that it has no outer validation-to-edit loop. It also does not clarify whether OpenCode receives equivalent domain guidance and retrieval support. This comparison is unfair because the methods do not receive equivalent knowledge support and feedback opportunities. I would like to see a comparison with the same tools, guidance, and budget, along with execution traces from OpenCode and GridRecon on the same tasks. The appendix currently provides only GridRecon cases.

  3. The paper still does not sufficiently explain how the requests were initially designed and written, how they were manually reviewed and filtered, or why these 36 task types are representative. Listing verification rules does not establish that the checks are reliable. Can they detect omitted constraints, incorrect objective priorities, or incorrect implementations that happen to produce the correct result on the current instance?

  4. Although the paper compares four models and several frameworks, it lacks an in-depth analysis of differences across models. The concrete cases all come from DeepSeek, and the stage-level statistics cover only a subset of its tasks. The paper does not explain where each model tends to fail, which errors can be corrected through feedback, or why the same framework performs so differently across models.

  5. The cross-network evaluation settings differ substantially between the main text and the appendix. The main text uses all 36 task types, yielding 432 instances, and reports 329 successful instances. The appendix uses a subset of 24 task types, yielding 288 instances, and reports an 88.5% success rate.

Questions for authors

See Weaknesses.

Reviewer_LzcDOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a multi-agent framework, GridRecon. The key challenge addressed in this work is that a task request should be correctly modeled/formulated (or tool calling) and its execution should satisfy physics. This work implements the proposed framework to check the validity of task orders and their implementation. A main contribution is the architecture design with four agents (or steps), which is not different from many common practices. The proposed work has been experimented with a predefined set of distribution-system operation tasks.

Critical strengths
  • A main contribution is the design and implementation of the multi-agent framework that jointly addresses the validity of prompts and its correct execution for distribution-system operations.
  • The numerical experiments show extensive comparison across multiple agent system implementation and LLMs.
Critical weaknesses
  • The paper needs substantial editing to improve readability in many places.
  • How does this work make sure the validity (requirement fidelity) with LLM? What is the deterministic validation? If this means another tool again, it seems falling into a circular dependency.
  • The numerical comparison section needs more in-depth analysis about why some tools work better than the others, rather than simply reporting and reading the results.
  • This ablation is not the best or most useful experiment to see the contribution, because removing a step from the proposed work is expected to lower the performance, and because the proposed agents are not modular but served as one framework. In particular, it does not make sense to remove one step when the framework is designed with four steps as a process.
Questions for authors
  1. Why does the work propose multiple agents? It develops multiple steps. Can we use a single agent for the four steps?
  2. It’s evident in the experiment that the token consumption could be inefficient. This may be because of having the four agents/steps. How can we improve that?
  3. The work shows the results from GPT-5.6-sol. Why don’t we use Codex?
  4. Why is the work specific to distribution system? Does the proposed work leverage any specific information/structure from the distribution system operations? Could it be applied or extended to transmission systems?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.