- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper proposes NC2C, an LLM-based framework for automatically formulating and solving non-convex optimization problems described in natural language. It incorporates Error Correction Loops (ECL) to address code execution failures and Feasibility Domain Correction (FDC) to refine solutions. The method is evaluated using GPT-5.1 and Qwen3-235B-A22B on four benchmarks: NL4Opt, NLP4LP, ComplexOR, and WireOpt. The reported results show improvements in success rate and execution rate over Vanilla prompting, Chain-of-Experts, OptiMUS, and Reflexion. Ablation studies and runtime analyses are also provided to examine the contributions and costs of different components.
Critical strengths
-
A practically meaningful attempt to automate the optimization workflow. The paper connects natural-language problem descriptions with mathematical optimization solvers in an end-to-end workflow. By integrating mathematical modeling, convexification, solver execution, and iterative refinement into a unified workflow, it reduces manual intervention in optimization modeling.
-
Encouraging results across several benchmarks and LLMs. The paper evaluates NC2C on four benchmarks using two different LLMs. For example, with GPT-5.1, NC2C achieves a success rate of 94.5% on NL4Opt and 72.2% on ComplexOR, compared with 90.2% and 63.9% for OptiMUS, respectively. The inclusion of ablation studies and runtime analysis also illustrates the behavior and computational overhead of the different components.
Critical weaknesses
- The contribution relative to related prior work could be more clearly articulated. I appreciate the effort to evaluate NC2C across several types of optimization problems. However, I had some difficulty identifying the contributions that distinguish NC2C from LLM-OptiRA (Peng et al., IEEE GLOBECOM Workshops, 2025; arXiv:2505.02091). Specifically, LLM-OptiRA already includes non-convexity detection, convexification, code generation, error correction, and feasibility validation. Its ECL and two-stage FDC mechanisms also closely resemble those described in Sections 3.4 and 3.6 of NC2C. Extending such a framework beyond wireless resource allocation is a worthwhile direction. What I found less developed is which additional technical challenges arise in this broader setting, and how NC2C addresses them beyond reusing the existing pipeline. I also noticed that LLM-OptiRA is not included among the experimental baselines. A controlled comparison using the same underlying LLM and comparable inference budgets would help distinguish the benefits of the generalization from those of the previously established components.
- The validity of the convexification process could be better supported. The paper presents an interesting approach to automatically transforming non-convex optimization problems into convex forms. However, I have concerns about whether the resulting formulations preserve the intended optimization problems. In Section 3.2, Equations (5)–(8) describe convexification through first-order Taylor approximations of objectives and constraints. These approximations do not preserve the original feasible region or optimal solutions without additional assumptions. The discussion does not establish when the transformations are exact, when they constitute relaxations or local approximations, or what guarantees apply to the resulting solutions. I also had difficulty reconciling the non-convexity identification in Appendix H.2 with the formulation in Appendix H.1. Under the stated assumptions of fixed channel gains, maximizing the sum of logarithmic rate functions with linear resource constraints is already a convex optimization problem. The optional minimum-rate constraints can also be rewritten as linear inequalities. Clarifying these mathematical conditions and providing independently verified examples would strengthen the evidence that NC2C performs valid convexification rather than merely producing executable reformulations.
- The experimental evaluation does not fully establish performance on genuinely non-convex problems or the quality of the resulting solutions. The results across four benchmarks and two LLMs are encouraging. However, I would have liked a clearer characterization of the optimization problems used to evaluate the central convexification claim. Appendix C states that NL4Opt originally consists primarily of LP and MILP problems, while NLP4LP contains linear and integer programming problems. Although the authors report filtering these datasets to retain instances requiring convexification, the selection criteria and the non-convex structures are not detailed. A breakdown distinguishing nonlinear non-convex objectives, non-convex constraints, integrality restrictions, and already-convex formulations would help establish what the framework is actually being tested on. There is also a distinction between feasibility and optimization quality that is not fully addressed. Section 4.3 describes Success Rate in terms of obtaining optimal solutions, but Equation (19) defines its indicator using feasibility. Reporting objective values or optimality gaps against independently verified reference solutions, together with a clearer description of the feasibility checks, would provide stronger evidence of the framework's effectiveness in solving non-convex optimization problems.
Questions for authors
- Contribution relative to prior work. Could the authors elaborate on the main technical differences between NC2C and LLM-OptiRA, particularly in the convexification and ECL/FDC components? It would be helpful to understand which challenges arise when extending the earlier framework to more general optimization problems, and how NC2C addresses them. A controlled comparison on WireOpt, using the same underlying LLM and comparable inference budgets, would also help clarify the contribution of these extensions.
- Mathematical validity of convexification. Under what conditions do the transformations in Equations (5)–(8) preserve feasibility or provide meaningful approximation guarantees? Could the authors also clarify the convexity classification in Appendix H.1–H.2 and the quadratic example in Appendix F.2.3? How is the correctness of generated convex formulations independently checked?
- Experimental validity and solution quality How were non-convex instances identified and selected from NL4Opt, NLP4LP, and ComplexOR? Could the authors provide a breakdown of the relevant problem structures, explain whether feasibility is independently verified against reference formulations, and report objective quality or optimality gaps on a representative subset? Clarifying the inference and correction budgets used for each baseline would also help interpret the comparisons.