Markdown edition

# FLARE tests AI grid restoration decisions under degraded reports

> A new simulation benchmark finds that AI agents often match a grid-restoration reference plan, but their worst cases and unsafe switch proposals keep independent safety checks central.

By BIG CHANGE Editorial

Published: 2026-10-19T11:40:34.898Z
Updated: 2026-10-19T11:40:34.898Z
Canonical: https://bigchange.ai/blog/flare-ai-grid-restoration-benchmark-degraded-reports

![Conceptual charcoal miniature of houses and local electricity poles, with an open disconnector prominent in the foreground.](https://bigchange.ai/api/media/file/flare-simulated-feeder-restoration-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE; no actual utility, outage, research apparatus or test result is depicted.

Researchers have built a [simulation benchmark called FLARE](https://arxiv.org/html/2610.23251v1) to test how language-model agents handle a narrow but consequential grid task: choosing how to restore electricity after a fault has already been isolated. The agents receive imperfect reports, can request more information and propose switch operations. A separate physics-based guard decides whether a proposed tie closure is safe in the simulated network.

The [October 16 preprint](https://arxiv.org/abs/2610.23251) reports 800 physical cases across five observation settings and 78,750 model episodes across its evaluations. It offers a way to examine restoration decisions and their failure modes. It reports no utility field trial or measured improvement in actual service.

## The big change

- **What changed:** FLARE ties an agent's information gathering and switch choices to simulated electrical consequences, rather than scoring an answer alone. The authors can compare decisions made with clean reports against decisions made with limited or erroneous ones.
- **Why it matters:** A grid researcher can inspect where an agent loses time or proposes an unsafe switch, while keeping a separate physical check in charge of whether that switch takes effect. The paper's rejected actions show why a plausible plan cannot serve as its own safety clearance.
- **What to watch:** The authors identify utility data, human-operator studies and fuller electrical validation as future work. Those checks would determine how far this benchmark can inform decision support beyond its simulated, post-isolation task.

## What the agents actually do

FLARE begins **after fault isolation**, when parts of a distribution feeder remain without power. In each case, the agent can ask for monitoring reports, customer-call information or fault-indicator readings; request an inspection or switch test; simulate a proposed operation; open a sectionalizer; close a tie switch; or stop. A tie switch can connect an unpowered area to another feeder, but the extra demand may overload equipment. The simulated task has a 240-minute horizon and limits on tool calls and switching operations, according to the [benchmark protocol](https://arxiv.org/html/2610.23251v1).

The authors pair CIGRE and SimBench feeder networks with measured German medium-voltage operating time series. They simulate faults, isolation and switching outcomes. Each of the 800 primary physical cases has a feasible reference restoration plan. The five versions of a case keep the underlying grid and fault fixed while changing what the agent sees: clean, sparse, noisy, stale or a compound setting labeled conflicting. That last setting also reduces monitoring coverage and increases several error rates, so its results cannot isolate the effect of contradictory reports alone. The paper's appendix says the stale SCADA age is a label on current-state readings, not a replay of old physical states.

## Average scores conceal difficult cases

The researchers evaluated five instruction-tuned models using the same task interface and deterministic decoding. Their score, *normalized regret*, compares each episode's weighted outage cost with the best feasible reference plan found for that physical case and with doing nothing. Zero matches the reference cost; one is no better than taking no action. The cost includes how long customers remain without power and weights customer classes, so a final restored-customer percentage does not describe the entire result.

Across the primary observation settings, [Table 2 of the paper](https://arxiv.org/html/2610.23251v1) reports mean regret from 0.082 for Qwen 2.5-7B to 0.148 for Qwen 2.5-3B. Every model has a median of zero, meaning at least half its episodes matched the reference cost. Yet 95th-percentile regret ranges from 0.814 to 1.000. The means therefore hide a smaller set of episodes with outcomes near the cost of leaving the simulated outage unrestored. These are benchmark scores, not measured customer outages.

More model parameters did not give a consistent advantage in this comparison: Qwen 2.5-7B scored better on mean and 95th-percentile regret than the tested Qwen 2.5-14B and four-bit Mistral Small 24B configurations. The paper also reports that under noisy observations, its idealized conventional FLISR reference scored 0.054 and a greedy method 0.060, compared with 0.089 for the best language model in that setting. Those references operated under the benchmark's common timing and score; the results do not establish a general ranking of operational systems.

## The guard caught proposals the models should not execute

Before a proposed tie closure changes the simulated grid, FLARE checks whether it would violate network shape, energize the isolated fault or fail AC power-flow limits on a trial copy of the physical state. This check uses information withheld from the agent. A rejected proposal still consumes time, but does not change the network.

The researchers then replayed 15,706 guard-rejected actions against cloned network states. They report that 15,142, or 96.4%, would have caused a physical violation. Thermal overload appeared in 8,165 rejected actions, with violation labels allowed to overlap. The absence of applied hard failures in the episodes reflects the guard's intervention; it does not show that the models independently selected safe actions.

For researchers, the useful distinction is between the quality of an agent's proposed plan and the independent check that keeps an unsafe proposal from taking effect. FLARE makes both visible. Its balanced positive-sequence power flow does not provide native three-phase or short-circuit validation, and the authors call for utility data and human-operator studies. The study evaluates simulated post-isolation restoration, not an autonomous utility controller or a proven change in real restoration times.

## Sources & further reading

- [FLARE: Physics-Grounded Evaluation of Language-Model Agents for Feeder Restoration with Degraded Observations](https://arxiv.org/html/2610.23251v1). BHVSP Subrahmanyam and Neeraj Choudhary, arXiv v1, submitted October 16, 2026. The full preprint supplies the task definition, simulated case construction, scoring, model results, guard analysis and stated limitations. It is the authors' evaluation, not a utility field study or independent replication.
- [arXiv record for FLARE](https://arxiv.org/abs/2610.23251). Confirms the title, authors, version and submission time and summarizes the principal results. The record does not add field validation.
- [FLARE code repository](https://github.com/dr-pandit-69/FLARE-neurips). Authors' public code, configuration and aggregate-result repository, read October 19, 2026. Its README says large grid substrates, generated panels, model weights and per-episode outputs are not stored there; BIG CHANGE did not run or reproduce the benchmark.

## Sources

- [FLARE: Physics-Grounded Evaluation of Language-Model Agents for Feeder Restoration with Degraded Observations, arXiv v1 full text](https://arxiv.org/html/2610.23251v1) — Primary methods, Tables 1–4, results and limitations. Authors' simulated evaluation; no utility field validation or independent replication.
- [arXiv record: FLARE](https://arxiv.org/abs/2610.23251) — Confirms v1 submission time, authors, title and abstract metrics; not separate empirical corroboration.
- [FLARE-neurips code repository](https://github.com/dr-pandit-69/FLARE-neurips) — Authors' README describes code, configs and aggregate results, plus omitted large substrate, generated panels, weights and per-episode outputs. Repository presence does not establish that BIG CHANGE ran the benchmark.