acceptodds
Withdrawn from ICLR 2027

Reflective Gradients: Feasibility–Task Tradeoffs in Solver Distillation with Verified Action Release

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Neural policies can violate explicit action constraints, while repairing every proposal with an optimization solver adds inference cost. We study whether detached solver supervision reduces repair dependence when an independent verifier remains responsible for action release. Admissible actions are finite unions of bounded polyhedra. The verifier checks proposals, requests repair when needed, and rechecks the decoded serialized action under the current state and rules. Under explicit compilation, arithmetic, state consistency, and consumer assumptions, every released action satisfies the encoded constraints independently of solver status. Feasibility verification alone gives no general approximation guarantee for branch-candidate selection; we reproduce severely suboptimal targets from a numerical repair service. A locally frozen study uses 20 paired trials per setting across disconnected and convex numerical tasks and a synthetic resource allocation application. Exact teachers, task reweighting, branch models, and fresh direct task optimization expose distinct feasibility, utility, and computational tradeoffs. Replacing numerical corrections with exact targets changes those tradeoffs without demonstrating a consistent learning advantage over task reweighting. Regret curves reveal threshold-dependent rankings, and the allocation task exposes poor learned utility at the fixed training budget. We provide conditional correctness proofs, analytic counterexamples, and auditable computation records. The findings motivate specifying teacher quality and evaluating complete decision cost under explicit task requirements rather than treating fewer repairs as sufficient evidence of useful constraint internalization.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neurosymbolic learning, constrained prediction, solver distillation, runtime verification, constraint internalization

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_7L1gOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper considers the setting of learning constrained optimization problem solutions via supervised learning. Unfortunately, the poor quality of this submission makes it nearly impossible to understand the motivation, contributions, and experiments of this work. From what I can understand, the work is motivated by the fact that learning optimization solutions through supervised learning depends on the quality of the target solutions themselves. The related work is severely underdeveloped, and no comparisons are made with existing work. The authors design three toy optimization problems, generate candidate solutions to these problems using different approaches, then use those candidates as targets for a learned predictor. I could not find a description of this predictor in the paper itself, it seems to be an MLP but that was only mentioned in passing the main text. The mathematical description of the optimization problems is also missing. Moreover, despite thoroughly reading the main text and appendix, I cannot understand the different approaches that were tried or why each was considered. This makes it impossible to evaluate whether the results are valid. The results themselves seem to contradict the assumption that learning is even valuable to begin with in this setting, as the direct task solver outperforms nearly all the learned approaches in terms of time, optimality, and feasibility. The presentation, informativeness, and soundness is well below an acceptable level for one of the top ML conferences in the world. Therefore, my decision is a clear reject.

Critical strengths

None

Critical weaknesses

I believe that the main fundamental weakness of this paper is that it was entirely LLM generated. I cannot enumerate every piece of evidence for this, but I can list some major offenders. Overall, an incredible amount of jargon is used throughout the paper without definition and without any clarifying examples, analogies, etc., and the paper lacks transition sentences that allow ideas to flow naturally from one to the next for a human reader. This makes the text incredibly terse and difficult to parse, and I could not follow the writing even as someone with years of experience in optimization. The introduction immediately jumps into describing the problem statement without any overarching or high-level motivation. The background and related work are severely underdeveloped, and it is clear that the authors did not perform a sufficient literature review before developing the thesis of the work. For example, I would recommend the authors look into the extensive work on learning to optimize, algorithm unrolling, optimization proxies, etc. that have been developed in the past 10-20 years. The entire hypothesis of the paper seems to rest on a “historical five-trial teacher study,” which has no citation and no description in the main text or appendix. The theoretical results are presented as if they are novel, but they are trivial restatements of known facts (e.g., Proposition 1 is just a restatement of feasibility). Acronyms such as ID are never defined, despite being a crucial part of the results. Finally, as stated in the summary, it is impossible to understand what the experiments are meant to demonstrate, and the results seemingly contradict the premise the learning is beneficial in this setting. This makes the work completely unacceptable for a top ML conference such as ICLR.

Questions for authors
  1. What is the motivation for studying this to begin with? Can you describe or identify an actual real-world problem and deficiency of the teacher that demonstrates this claim?

  2. Why were these optimization problems chosen? What qualities make them worth studying for your setting?

  3. What was the motivation for each of the methods considered?

  4. What exactly was being learned? What was the architecture?

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

A neural network is supposed to take actions that must obey hard rules (limits and budgets), but it is often found to violate these. One common fix is to put a solver after the network. If the network’s proposal breaks a rule, the solver nudges it to the nearest action that follows the rules. This works, but running the solver every time costs computation. The central idea of the paper is to train the network to copy the solver’s corrections, so it learns to follow the rules on its own and needs the solver less often (called reflective gradient). Separately from learning, the authors build a strict checker. It verifies every action with exact arithmetic (no rounding tolerance) before releasing it. It even re-checks the action after it has been packaged for sending.

Critical strengths

The positive points of the paper are: a) use of strong baselines (direct optimization), b) results are reproducible and satisfy the statistics (settings and seeds were fixed before testing), and c) an independent checker makes the safety claims trustworthy.

Critical weaknesses

The negative points of the paper are: a) no working method comes out of it, b) the tasks used for evaluation are simple, c) uncertainty in settings is not studied, d) no direct comparison with established methods, e) speed differences are minuscule, and f) novelty is limited.

Questions for authors

The detailed comments are given below.

  1. The paper diagnoses problems but offers no approach that clearly beats simple baselines.
  2. There are two 2D problems and one 16D generated problem. Nothing real or large-scale is tested.
  3. Because the rules and goal are fully known and cheap to solve, direct optimization wins. The interesting case of uncertainty where solving is expensive, or the goal is only partly known, is not studied.
  4. Most latency gaps are fractions of a millisecond on one CPU, so they matter little in practice.
  5. The writing is very dense and has readability issues. The title suggests a new gradient method, but it seems like the novelty is extremely limited.
  6. Direct comparison with established methods such as DC3 and HardNet is missing.
Reviewer_j4mmOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper addresses the problem of producing policies whose actions comes with constraints. It raises the issue that corrections provided by optimization solvers are feasible but may be suboptimal. It mainly is a controlled empirical and theoretical study of feasible-correction teachers in constrained decision making: whether learning from “verified feasible” correction targets actually improves task performance, and what can go wrong when feasibility is confused with optimality.

The experiments use 20 paired seeds on three synthetic tasks: a disconnected 2D feasible set, its convex control, and a 16-dimensional resource-allocation task with four modes. The protocol freezes teacher masks, parents, selections, and trial counts before final testing, uses exact release checking for every method, and reports native feasibility, guarded success at regret threshold 0.05, coverage, regret, and complete decision latency.

Critical strengths
  1. Investigating methods for ensuring constraint satisfaction of actions produced by nerural policies is a useful problem to address. The paper separates three things that are often conflated: exact encoded feasibility, correction optimality, and downstream task utility. This is a useful diagnostic contribution.

  2. Proposition 2 gives a clean, reproducible example showing that verified candidate selection can have an unbounded global approximation ratio even for a bounded convex union. This directly challenges the assumption that “exactly checked feasible” means “good correction.”

Critical weaknesses
  1. All tasks are synthetic with observable objectives and constraints. The 16D allocation task is larger than the 2D tasks but still far from industrial control, partial observability, noisy dynamics, or real deployment.

  2. It assumes correct compilation, parsing, rational arithmetic, context validity, and consumer behavior. It is not a machine-checked proof of the Python implementation and does not cover trajectory safety, availability, or deadlines.

  3. This is a very poorly written paper with substantial AI assiatnce: 1) there's no discussion on the background of the problem, nor the precise formulation of the problem, nor any motivations for the proposed methods. 2) the authors also agree that this is substantial AI involve ement in the paper: Generative AI tools assisted with research ideation, literature discovery, experiment design, implementation, synthetic-data generator design,mathematical claims and proof drafting,diagnostic analysis,figure generation,and manuscript drafting and editing. 3) the authors need to digest the AI interactions to make the paper at least readable in the first place!

Questions for authors
  1. A run time checker can check the constraints satisfaction, but what is an explicit execution contract?
  2. What is a gradient-reflection operator?
  3. A prior teacher refers to what?
  4. No explanation for the symbols used in eqaution (1)
  5. The paper introduces many near-synonyms that differ subtly: repair vs. correction vs. projection vs. teacher vs. guard vs. verifier vs. release.
  6. Many terminologies are used without any explanation like the above!

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

Related papers

Discussion (0)

Sign in to comment.