acceptodds
Withdrawn from ICLR 2027

When Objectives Agree But Rules Do not: The Hidden Cost of Exposing Rule Errors In Optimization Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

An optimization-program revision may attain the expected optimum while enforc- ing the wrong rule, because the plans that expose the error can be worse than every operationally plausible decision. Such a revision passes any check that compares optimal objective values. We formalize this value–information barrier.The criti- cal loss ℓ(i, j) is the least objective degradation at which two rule interpretations disagree; its maximum against all alternatives exactly characterizes target identifia- bility from complete-plan membership answers. It also bounds the blind spot of objective-value checks. An LP slack law bounds exposure cost using rule slack and marginal value; a public-LP audit checks 63 constructed bound changes. A stored- data audit finds a strong association between the target’s own optimum loss and its critical loss (ρ =0.823 across 24,192 semantic classes). One post-hoc public complete-plan case has a model-derived 17.6% barrier; broad full-program LLM evidence remains unmeasured. Two sealed three-domain studies show a steep rise: 1.9–4.2% of update targets are identifiable using only optimal plans, compared with 64.7–67.8% at 10% loss. Factorial studies in two domains reproduce the frontier, and appended rules are harder to expose than retracted ones (67.8% versus 95.1% of classes at 10% loss). Critical-loss-Aware Planning (CAP) is an exact finite- budget reference that separates witness availability from question allocation and shows that isolating targets and shrinking ambiguity are different objectives. On 24 public benchmark revisions, 18/23 executable errors from three local models need positive local loss to expose. The evidence establishes a replicated finite-domain phenomenon and a limited public-task bridge; it does not estimate natural-history prevalence, open-ended generated-program loss, or human clarification cost.

Primary area:
learning theory
Keywords:
Large Language Models, Optimization Modeling, Constraint Acquisition

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_jTKXOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper identifies and formalizes the value-information barrier in combinatorial and linear optimization revisions, where semantically distinct rule interpretations yield identical optimal objective values on a base domain, rendering specification errors invisible when inspecting only optimal plans.

Critical strengths

S1. Formalization of Model Disambiguation: The paper identifies a fundamental tension between plan optimality and semantic verification in optimization modeling. Formalizing critical loss provides a crisp, principled framework for explaining why evaluating optimal solutions alone fails to uncover subtle constraint errors in optimization programs.

S2. Theoretical LP Slack Analysis: The theoretical bounds derived for continuous linear programs offer valuable mathematical insights into how constraint perturbations and LP slackness dictate the exact objective loss necessary to expose rule differences.

Critical weaknesses

W1. Abstract Clarity and Information Density. Problem Formulation Overwhelmed by Numbers: The abstract introduces dense empirical statistics (e.g., 278,208 sealed class pairs, 35,106 same-optimum pairs, 35.6% requiring extra loss, 1.9–4.2% vs. 64.7–67.8% certifiability) before establishing a clear intuitive narrative or a crisp formal definition of the core problem.

Unclear Terminology: Key constructs such as "value-information barrier," "complete-plan membership answers," "target identifiability," and "registered append targets" are introduced rapidly without sufficient context for a general machine learning audience.

W2. Practical Viability and Cognitive Load of Complete-Plan Auditing. Auditing full, suboptimal decision plans imposes an unrealistic cognitive burden on domain experts when applied to real-world, high-dimensional optimization problems. Expecting human reviewers to inspect monolithic, multi-variable plans to verify subtle constraint rules lacks practical justification compared to directly reviewing localized constraint definitions or targeted counterexamples.

W3. Strong Assumptions on Candidate Pools and Idealized Oracles. The framework relies on a pre-defined candidate rule pool (H0\mathcal{H}_0) and fails entirely if an LLM or developer introduces an unexpected error outside this set. Additionally, collapsing candidate rules requires computationally intractable satisfiability checks for large-scale models, while assuming exact, noise-free binary oracle feedback ignores human auditing fatigue and systematic misinterpretations.

W4. Limited Technical Novelty and Scalability of CAP. Critical-loss-Aware Planning (CAP) adapts established decision-tree dynamic programming without significant algorithmic novelty and suffers from exponential time complexity O(Bm2n)O(B m 2^n). Because it is restricted to tiny candidate pools (n≤32n \le 32) as an offline theoretical baseline, the lack of approximate or scalable heuristics creates a major gap for practical deployment.

W5. Unrealistic and Synthetic Empirical Foundations. The evaluation relies almost entirely on micro-scale, fully enumerable toy domains (e.g., 7 jobs on 3 machines, 12-item knapsack) that do not reflect the scale or complexity of industrial optimization models. Furthermore, the continuous LP experiments use artificial, manually perturbed inequality bounds that fail to represent realistic specification revisions or code generation errors.

W6. Disconnect Between LLM Framing and Empirical Evidence. Despite heavily framing the paper around verifying LLM-generated code, empirical support on full LLM programs is sparse. The evaluation relies on restricted local plan spaces, and post-hoc testing on public cases yields only a single instance with a true value-information barrier, leaving the real-world prevalence of this issue unproven.

Questions for authors

W2--W6 above.

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

This paper investigates the problem where different models can lead to the same solution and false negatives. Some theory is developed to characterize this issue, and some experiments are done.

Critical strengths

The studied problem is useful.

Critical weaknesses

Although the problem looks interesting, I'm very concerned about the current presentation.

This paper is full of weird terminology that looks LLM-generated, such as "replay", "audit", and "bridge". After a brief check, the code repository also looks LLM-generated. I don't think the authors even polished the paper. The theoretical results define many new concepts, but the experiments don't actually use them.

The AI disclosure says "reported values were recomputed from sealed artifacts with automated source, hash, and source-to-claim checks". I don't think the authors take any responsibility for the results.

Overall, this paper is not readable by a human, and it cannot be published at ICLR in its current form.

Questions for authors

N/A.

Reviewer_bo8AOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper is unreadable.

Critical strengths

None.

Critical weaknesses

Clarity. This paper is unreadable.

  • For example, it is hard to understand the abstract from the third sentences: "We formalize this value–information barrier. The critical loss ℓ(i,j)\ell(i,j) is the least objective degradation at which two rule interpretations disagree; its maximum against all alternatives exactly characterizes target identifiability from complete-plan membership answers on a finite shared domain." There are many invented terminologies with no definition, such as "value–information barrier", "critical loss", "rule interpretation", "target identifiability", "complete-plan membership".
  • Section 3 starts with little introduction of the problem formulation and proceeds to eight propositions/theorems.
Questions for authors

Could you defined and explain these terminologies?

  • "value–information barrier", "critical loss", "rule interpretation", "target identifiability", "complete-plan membership" in the abstract.
  • "public complete-plan witness", "translation-invariant base-range sensitivity analysis", "response signatures", "current-interface class", "finite shared base domain", "certifiable" in Section 3.

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

Related papers

Discussion (0)

Sign in to comment.