acceptodds
Withdrawn from ICLR 2027

Repair the Action, Preserve the Intent: Minimal Feasibility Projection for Power-Grid Agents

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Language-model agents can translate operator requests into power-grid actions, but an executable switching or redispatch command can violate connectivity, generator, voltage, thermal, or operational-goal constraints. Re-prompting until a simulator accepts a command supplies neither a minimality guarantee nor a boundary on how much of the request may be replaced. We formulate intent-preserving feasibility repair: search a declared finite mixed discrete-continuous neighborhood in lexicographic order of action type, target, support, magnitude, and finally operating cost. Our implementation, IPFR, validates semantic edit buckets with AC power flow, returns the lowest-cost action in the first feasible bucket, and explicitly refuses when the neighborhood contains no admissible repair. This yields soundness and lexicographic minimality relative to the declared neighborhood, without assigning commensurate weights to semantic and monetary quantities. On 120 held-out scenarios from IEEE 57- and 118-bus systems, IPFR finds an AC-feasible action in all 108 repairable cases and refuses all 12 infeasible controls. In 73 conditional, adversarial priority-conflict cases it preserves the proposed action type and target in every case, whereas cost-first feasibility projection does so in none. These cases test contract compliance, not field prevalence. IPFR uses 4.77 AC validations per repairable case versus 60.07 for an exhaustive lexicographic oracle and agrees with that oracle on every selected action. Euclidean and development-tuned scalar projections match its actions at ordinary units but require exhaustive search and lose unit invariance. The contribution is a verifiable boundary between correcting an unsafe agent and silently replacing what it was asked to do.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
AI safety, agentic systems, lexicographic optimization, feasibility repair, power systems, runtime assurance

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_NR2QOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses a relevant problem in power systems, but the proposed methodology is relatively simplistic, with limited technical novelty. The experimental evaluation is also preliminary, relying on small test systems and synthetic scenarios, which provides insufficient evidence of practical applicability. Overall, the work is still at an early stage and requires further development and more comprehensive validation.

Critical strengths
  1. The paper presents a complete study.
  2. The paper addresses an interesting problem.
Critical weaknesses
  1. The proposed method mainly relies on lexicographic ordering and feasibility checking over a predefined candidate set. The overall algorithm is relatively straightforward, and the theoretical results largely follow from the search procedure.
  2. The experiments are conducted on relatively small power systems with synthetic scenarios. More importantly, the proposed method does not outperform conventional projection methods in terms of feasibility or intent preservation. Its practical advantages over simpler alternatives remain unclear.
Questions for authors

Please see weaknesses

Reviewer_MguMOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to repair an infeasible action proposed by a power-grid agent while respecting a declared order of permitted changes. IPFR prioritizes preserving the action type and target, followed by the set of affected generators and the adjustment magnitude, with operating cost used to break ties. It searches a finite candidate set using AC validation and refuses when no candidate is feasible. The paper establishes guarantees relative to this candidate set and validator, and reports successful repair of all 108 repairable cases and refusal on 12 controls from held-out IEEE-57 and IEEE-118 scenarios.

Critical strengths
  1. The paper addresses a useful problem in agent-assisted grid control: a physically feasible replacement can change the equipment or operation specified in the original proposal. Making the permitted changes explicit gives the repair process a clear scope and allows its decisions to be checked. The formulation connects this operational concern to a concrete selection rule with guarantees within the declared candidate set.

  2. The experimental materials support inspection of the method’s behavior. The shared candidate sets, detailed validation records, and exhaustive oracle allow the repair decisions to be checked directly. The paper also separates constructed priority conflicts from cases where the original target cannot be retained, and clearly states the limits of its static grid model. These choices make the evidence for implementation correctness easier to assess.

Critical weaknesses
  1. The paper does not sufficiently explain how its contribution differs from prior work on preference-aware safety filtering and priority-based behavior specification. In Safe Reinforcement Learning via Shielding (Alshiekh et al., AAAI 2018), the authors describe a shield that selects the highest-ranked safe action from a learner-provided list, with fallback to actions outside the list when necessary. Liability, Ethics, and Culture-Aware Behavior Specification using Rulebooks (Censi et al., ICRA 2019) formalizes priorities among behavioral rules and recovers lexicographic ordering when the rules form a linear order. IPFR addresses a different setting, using AC feasibility checks within a declared finite neighborhood and refusing when no feasible repair exists. Its typed edit hierarchy and repair contract may therefore offer a distinct contribution. However, the manuscript does not discuss these relevant precedents or establish what additional capability or practical benefit its formulation provides relative to them. This leaves the methodological novelty insufficiently established.

  2. The experiments do not establish a clear advantage over comparably implemented repair baselines. Euclidean and tuned weighted projection already match IPFR’s selected actions on all 108 repairable cases. Replaying the supplied candidate records with early stopping for Euclidean projection gives the same decisions and validator-call counts as IPFR across all 120 scenarios, including the same 4.77-call average on repairable cases. This replay uses the supplied feasibility outcomes without rerunning AC power flow. Figure 3 also rescales an already normalized cost term, so it does not establish a failure under consistent currency conversion. The current results therefore leave the practical benefit of the proposed hierarchy over normalized projection insufficiently supported.

Questions for authors
  1. What is the main contribution of IPFR beyond ranked shielding and existing priority-based decision frameworks? Please clarify which aspect of the proposed formulation introduces a new capability or insight, and why the combination of the edit hierarchy, AC validation, and refusal mechanism represents a substantive advance.

  2. How does IPFR compare with baselines that use comparably efficient search strategies? In particular, Euclidean projection could order candidates by distance and stop after evaluating the first distance group containing feasible candidates. A comparison using this implementation would help establish whether IPFR retains an advantage in validation effort or in preserving the declared action priorities.

  3. Euclidean and tuned weighted projection select the same actions as IPFR on all reported repairable cases. Under what practically relevant operating conditions would their decisions differ? Could the authors provide examples or experiments showing when the explicit hierarchy improves repair decisions, including any tradeoffs in operating cost or control effectiveness?

  4. How should the cost-rescaling results be interpreted given that the weighted baseline already uses a normalized cost term? Please distinguish the effects of consistently changing currency units from those of changing the relative importance of cost. Would the conclusions about robustness still hold under consistent unit conversions and variations in operating conditions?

  5. The paper clearly describes the priority-conflict cases as a conditional stress test. How often does IPFR provide a meaningful benefit when proposals are generated without conditioning on a conflict between cost and target retention? An evaluation in this setting would help establish how broadly the reported findings apply.

  6. How sensitive are the repair success rate, priority preservation, and computational cost to the construction of the candidate neighborhood? Would the main conclusions hold with a wider range of candidate actions or different search budgets? Please discuss the practical limits of expanding the neighborhood and which extensions are most likely to improve the method’s usefulness.

Reviewer_AUkuOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies post-hoc feasibility repair for structured power-grid control actions a0=(τ0,b0,Δp0)a_0 = (\tau_0, b_0, \Delta p_0) that violate AC power-flow or operational constraints. The authors define a five-component typed edit vector comparing a candidate action aa against a0a_0 across action family dτd_\tau, target identity dTd_T, generator support cardinality dSd_S, normalized active-power magnitude dMd_M, and post-solve generation cost gap Cx(a)C_x(a). Their algorithm IPFR validates buckets sequentially via PYPOWER AC power flow, and tie-breaks within the first feasible bucket by Cx(a)C_x(a).

Critical strengths
  • Explicit separation of candidate semantics and physical validation
  • Honest refusal controls and failure accounting
Critical weaknesses
  • first of all, I don't this the set up the problem is valid, espeically for the power grid numerical problems. Please justify the approach, not just from empirical evaluation.
  • IPFR is behaviorally identical to Euclidean/Weighted, and the 12.6×12.6\times call saving is a reporting choice.
  • The headline effect is guaranteed by construction, held-out rules differ, and no agent is evaluated. Despite the "language-model agents" framing (Sec. 1), no LLM is evaluated
Questions for authors
  • Replaying candidate_evaluations.csv, an early-stopped Euclidean search needs 4.914.91 calls and returns IPFR's action in all 120120 held-out scenarios. What does the lexicographic order provide that early-stopped Euclidean or Weighted search does not? Can you exhibit any scenario where they differ?

  • Since the 73/7373/73 vs. 0/730/73 result follows from the acceptance rule at experiment.py:457, can you report results on unconditioned scenarios, and on proposals produced by an actual LLM agent from the operator templates?

  • About 68.8%68.8\% of "AC calls" stop at bound or connectivity checks before runpf. Can you report power-flow solves and wall-clock time separately, and quantify how the post-held-out Stage 4 correction changed the held-out results?

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

Related papers

Discussion (0)

Sign in to comment.