acceptodds
Under review as a conference paper at ICLR 2027

EvoMIDFL: Agentic Self-Evolving Decision-Focused Learning For Mixed-Integer Programming

Abstract

Decision-focused learning (DFL) embeds optimization problems into learning architectures and trains using downstream decision loss rather intermediate prediction errors alone. In mixed-integer DFL (MIDFL), however, the discontinuous prediction-to-decision map obstructs informative gradient propagation. Existing approaches typically rely on a fixed approximate backward mechanism with hand-tuned configurations, limiting the adaptation across problems, scales and leaving potentially useful gradient combinations unexplored. Recent advances in agentic evolution offer a promising way to search and refine mechanisms through successive evolution. To this end, we propose EvoMIDFL (Evolutionary Mixed-Integer Decision-Focused Learning), an agentic self-evolving MIDFL framework that adaptively discovers, configures, and composes surrogate backward-gradient policies across diverse settings. EvoMIDFL employs a validation-guided actor–critic workflow supported by a three-layer memory architecture: the Gradient Actor proposes gradient estimators, combinations, and configurations; the Critic Agent diagnoses failures and filters unsafe candidates; and the Memory Agent stores formulation-level experience, mechanism-level evidence, and transferable skills to guide subsequent evolution. Retrieved policies are further refined on the target task to mitigate negative transfer. Extensive experiments show that EvoMIDFL not only improves downstream decision quality over fixed MIDFL backward methods, but also discovers reusable policies across scales and scenarios.

Primary area:
optimization
Keywords:
Agentic, Self-Evolution, Decision-Focused Learning, Backward mechanisms, Mixed Integer Programming

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_LiAvOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EvoMIDFL, an agentic framework that evolves surrogate backward-gradient policies for mixed-integer decision-focused learning. The method combines several gradient estimators, validates candidate policies from the same training checkpoint, and maintains a three-layer memory for transferring successful configurations across scales and problem families. Experiments on bipartite matching and knapsack show that EvoMIDFL is often competitive with fixed backward methods, with small improvements in several settings and evidence of cross-task transfer.

Critical strengths
  1. The paper gives a coherent framework for selecting, configuring, and composing heterogeneous surrogate gradients.
  2. The memory-transfer ablations are useful.
Critical weaknesses
  1. The empirical gains are generally small and not accompanied by standard deviations, confidence intervals, or paired significance tests.
  2. The evolution process repeatedly selects policies using the same validation set over many rounds and hundreds of proposals, creating a substantial risk of validation overfitting.
  3. The contribution of the LLM itself is not isolated from a similarly budgeted non-LLM search or hyperparameter optimization baseline, and the computational/token cost is not reported.
Questions for authors
  1. Can you report per-seed results and statistical tests for the main comparisons, especially matching-225, matching-900, and knapsack-2000?
  2. How much wall-clock time, solver calls, and LLM/token usage does EvoMIDFL require compared with fixed baselines and an equally budgeted random or Bayesian policy search?
  3. How do the results change with an independent meta-validation set, or with a reversed transfer direction (knapsack to matching) and an unseen MILP family?
12 comments
AuthorsOpenReview ↗

General Response to Reviewer LiAv

We thank the reviewer for the thoughtful evaluation and positive remarks regarding our surrogate-gradient adaptation framework and memory ablations.

Guided by your suggestions, we have expanded our experimental analysis in the revised manuscript as follows:

  • LLM-driven contributions: We reclaimed the contributions of LLM-driven agents compared with aotoML methods.
  • Per-Seed Variance & Statistical Scope: We reported full per-seed results in the appendix, along with the corresponding standard deviations.
  • Cost analysis and LLM usage: We supplemented the cost analysis of wall-clock time, solver calls, and representative LLM usage.
  • Robustness & Generalization: We evaluated robustness on two distinct validation sets, tested reverse-memory transfer (Knapsack →\rightarrow Matching), and validated our work on an unseen MILP family (Newsvendor).

We are grateful for these recommendations, which have allowed us to demonstrate the rigor and practical scope of EvoMIDFL more comprehensively. Detailed discussions and experimental evidence for each point are provided below.

AuthorsOpenReview ↗

Detailed Response to Reviewer LiAv (Weakness)

Weakness: LLM-driven benefits.

We thank the reviewer for pointing out the weakness. To thoroughly evaluate the contribution of LLM-driven evolution, we first conducted three new automated search experiments (AutoML methods): Random Search, Evolutionary Search, and Bayesian Optimization, and then clarified the benefits of LLM-driven methods.

1. New experiments of AutoML:

  • Settings: In AutoML experiments, we replaced the Actor with each AutoML method in turn. Results demonstrate that under an identical budget of validation rounds, EvoMIDFL–Full achieves lower mean test regret than all three baselines across every evaluated setting (in Table 3: Comparison with conventional search baselines on Matching and Knapsack in the main paper).
  • Empirical gains. Empirically, our quantitative results confirm that LLM-driven, diagnostic-guided policies yield superior decision quality compared to non-LLM baselines under an identical evolutionary budget. As shown in Table 3, given the same number of evolution rounds, EvoMIDFL–Full achieves the lowest mean test regret across all four Matching and Knapsack benchmarks. Relative to the strongest AutoML baseline on each task, the most substantial regret reductions occur on the Matching benchmarks: 7.4% on Matching-900 k-1(from 47.862 to 44.304) and 9.2% on Matching-1600. On Knapsac000 and -2000, the reductions in mean test regret per item are more modest, yet consistent at 1.1% and 1.5%, respectively.

2. LLM-driven benefits:

Based on the conventional AutoML results and previous comparison with fixed baselines, EvoMIDFL overcomes both (i) the blind black-box parameter tuning of conventional AutoML and (ii) the rigid, hand-crafted heuristics of fixed baselines that fail to adapt across problems or scales. The two key differences between EvoMIDFL and conventional AutoML search or fixed baselines are detailed below:

(1) Diagnostic-Guided Reasoning (EvoMIDFL v.s. AutoML)

  • How EvoMIDFL works. The LLM Actor agent explicitly generates proposals based on training diagnostics, the parent policy, and recent trial outcomes. For example, when validation regret plateaus alongside fluctuating gradient norms, the Actor can diagnose optimization instability and prioritize a coordinated candidate, such as switching the gradient estimator while simultaneously tightening the clipping threshold to directly resolve the underlying failure mode.
  • Conventional search. In contrast, conventional AutoML methods treat policy search as a black-box optimization problem relying solely on scalar validation scores. Specifically, Random Search samples uniformly, Evolutionary Search perturbs candidates near historically strong policies, and Bayesian Optimization fits a surrogate model exclusively on past validation scores without inspecting intermediate training dynamics. Crucially, conventional AutoML methods typically optimize scalar validation metrics via black-box sampling, lacking visibility into ongoing training dynamics to diagnose failure modes or perform targeted, structural policy refinements.
  • Benefits: The targeted-validation diagnosis in EvoMIDFL overcomes the blind black-box parameter tuning of conventional AutoML, which is supported by the empirical experiments of EvoMIDFL with AutoML under identical evolution rounds.

(2) Paradigm Shift Toward Dynamic, Training-Aware Policy Adaptation: (EvoMIDFL v.s. fixed baselines)

  • Conventional DFL: Conventional decision-focused learning (DFL) pipelines typically treat the backward gradient estimator as a static design choice fixed prior to training, requiring practitioners to manually select, tune, and validate candidate surrogate formulations through laborious trial and error.

  • New paradigm: EvoMIDFL shifts this paradigm by formulating gradient-policy selection as a dynamic, sequential decision process that continuously adapts alongside evolving training dynamics. By tightly coupling structured policy proposals with validation-gated promotion, the framework autonomously navigates combinatorial spaces of estimators, hyperparameters, and stabilization mechanisms. This closed-loop automation eliminates the need for exhaustive manual candidate enumeration and domain-specific hand-crafting, substantially reducing human engineering effort while adapting seamlessly to varying optimization stages and problem structures.

  • Balancing accuracy and stability: EvoMIDFL ensures both high accuracy and strong stability in final decision quality. Across the majority of evaluated settings, EvoMIDFL outperforms individual single-policy baselines in terms of test regret in most occasions; even in the few scenarios where performance gains are more modest, the margins remain small and competitive, demonstrating reliable cross-scenario robustness without significant performance degradation.

AuthorsOpenReview ↗

Detailed Response to Reviewer LiAv (Weakness cont.)

Table 3: Comparison with conventional search baselines on Matching and Knapsack (in the main paper). RS, ES, and BO denote random search, evolutionary search, and Bayesian optimization. Knapsack regret is per item.

MethodMatching-900 regret ↓Matching-900 RMSE ↓Matching-1600 regret ↓Matching-1600 RMSE ↓Knapsack-1000 regret/item ↓Knapsack-1000 RMSE ↓Knapsack-2000 regret/item ↓Knapsack-2000 RMSE ↓
RS51.254 ± 0.7048.736 ± 1.60472.654 ± 6.0419.620 ± 0.9470.289 ± 0.0222.260 ± 0.3590.304 ± 0.0642.439 ± 0.647
ES49.559 ± 4.8638.441 ± 2.15470.473 ± 5.4809.742 ± 0.8310.266 ± 0.0111.885 ± 0.0050.265 ± 0.0092.174 ± 0.132
BO47.862 ± 1.5339.372 ± 1.10672.429 ± 2.0407.126 ± 1.2510.270 ± 0.0231.979 ± 0.1270.268 ± 0.0043.604 ± 0.399
EvoMIDFL–Full (Ours)44.304 ± 0.3647.204 ± 0.43164.011 ± 0.3336.939 ± 0.9780.263 ± 0.0351.873 ± 0.2180.261 ± 0.0242.025 ± 0.237
Reviewer_ogv7OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EvoMIDFL, an agentic framework for mixed-integer decision-focused learning. Instead of relying on a fixed surrogate gradient estimator, EvoMIDFL uses an LLM-based actor–critic workflow to adaptively select, configure, and combine existing backward mechanisms. A three-layer memory system supports the reuse and refinement of gradient policies across tasks. Experiments on matching and knapsack problems show improvements in downstream decision quality over several fixed-gradient baselines.

Critical strengths
  1. The paper addresses a meaningful problem in mixed-integer decision-focused learning. Adaptively selecting and configuring surrogate gradient estimators is a reasonable and potentially useful direction.

  2. The proposed framework is well-structured, with validation-guided policy evolution and memory-based adaptation. The experiments cover multiple problem scales and provide useful analyses of the evolution process.

Critical weaknesses

The added value of LLM-based agentic evolution is not sufficiently established. The motivation for adaptive gradient-policy selection is reasonable. However, the search space is predefined and bounded, and candidate policies are ultimately selected based on validation performance. This makes the problem closely resemble conventional hyperparameter or algorithm configuration.

While the experiments demonstrate the benefits of adaptive gradient-policy evolution, they do not establish whether LLM-guided search offers advantages over simpler search methods. Given the relatively modest improvements over the strongest fixed-gradient baselines, this missing comparison makes it difficult to assess the paper's central agentic contribution.

Questions for authors

What is the specific advantage of using an LLM agent for gradient-policy evolution?

Given the bounded and structured search space, could the authors compare EvoMIDFL against conventional approaches such as random search or evolutionary search under matched candidate-evaluation and computational budgets?

I find the motivation for adaptive gradient-policy selection convincing, but the necessity and benefits of LLM-based reasoning remain unclear. Demonstrating that the agent can achieve better search efficiency or final decision quality than these simpler alternatives would substantially strengthen the paper.

This is my main concern, and I would be willing to raise my rating if the authors can convincingly address it.

7 comments
AuthorsOpenReview ↗

General Response to Reviewer ogv7

We thank the reviewer for the thorough and insightful assessment. We sincerely appreciate the positive feedback on the motivation for adaptive gradient-policy selection and memory-based adaptation.

We understand that the main concerns are what LLM-driven evolution adds and how much it improves results over simpler search alternatives. We addressed both points below:

  • Role of the LLM: We clarified the two benefits of LLM-driven EvoMIDFL compared with autoML and fixed baselines.

  • Comparison with Conventional Search: We have introduced extensive empirical comparisons against conventional search baselines (Random Search, Evolutionary Search, and Bayesian Optimization). Across our evaluations, EvoMIDFL–Full consistently achieves lower mean test regret and superior stability.

We detail the analysis below. We believe these detailed responses and new experimental validations thoroughly address your concerns and clarify the novelty and contribution of our framework.

AuthorsOpenReview ↗

Detailed Response to Reviewer ogv7 (Q1)

Q1: What does the LLM add to gradient-policy evolution?

To answer the question, we first ran experiments with the AutoML methods (results are listed in the detailed response to Q2). Based on the conventional AutoML results and previous comparison with fixed baselines, EvoMIDFL bridges training dynamics with closed-loop validation, overcomes both (i) the blind black-box parameter tuning of conventional AutoML and (ii) the rigid, hand-crafted heuristics of fixed baselines that fail to adapt across problems or scales, leading in better final decision quality.

Below, we detail the core advantages of the LLM-driven EvoMIDFL framework compared to conventional AutoML search and static fixed baselines:

  1. Diagnostic-Guided Reasoning (vs. AutoML Search):
  • How it works: The LLM Actor agent explicitly generates proposals based on training diagnostics, the parent policy, and recent trial outcomes. For example, when validation regret plateaus alongside fluctuating gradient norms, the Actor diagnoses optimization instability and intelligently prioritizes a coordinated candidate (such as switching the gradient estimator while simultaneously tightening the clipping threshold).
  • Compared with baselines: Conventional AutoML methods treat policy search as a black-box optimization problem relying solely on scalar validation scores. Specifically, Random Search samples uniformly, Evolutionary Search perturbs candidates near historically strong policies, and Bayesian Optimization fits a surrogate model exclusively on past validation scores without inspecting intermediate training dynamics. Crucially, conventional AutoML methods typically optimize scalar validation metrics via black-box sampling, lacking visibility into ongoing training dynamics to diagnose failure modes or perform targeted, structural policy refinements.
  • Empirical gains: Given the same evolutionary round budget, EvoMIDFL–Full consistently achieves lower mean test regret than all three AutoML baselines. Relative to the strongest baseline at each scale, EvoMIDFL–Full reduces mean test regret by 7.4% on Matching-900 (from 47.862 to 44.304) and 9.2% on Matching-1600 (from 70.473 to 64.011), alongside consistent improvements on the Knapsack-1000 benchmark (e.g., 0.266 to 0.263). The results demonstrate that EvoMIDFL shows a strong ability to find suitable gradient policies that lead to better final decision quality.
  1. Paradigm Shift From Static Manual Tuning to Dynamic, Training-Aware Adaptation (vs. fixed baselines):
  • Conventional decision-focused learning (DFL) pipelines typically treat the backward gradient estimator as a static design choice fixed prior to training, requiring practitioners to manually select, tune, and validate candidate surrogate formulations through laborious trial and error.

  • Closed-loop automation. EvoMIDFL shifts this paradigm by formulating gradient-policy selection as a dynamic, sequential decision process that continuously adapts alongside evolving training dynamics. By tightly coupling structured policy proposals with validation-gated promotion, EvoMIDFL autonomously navigates combinatorial spaces of estimators, hyperparameters, and stabilization mechanisms. This closed-loop automation eliminates the need for exhaustive manual candidate enumeration and domain-specific hand-crafting, substantially reducing human engineering effort while adapting seamlessly to varying optimization stages and problem structures.

  • Benefits. EvoMIDFL ensures both high accuracy and strong stability in final decision quality. Across the majority of evaluated settings, it consistently outperforms individual single-policy baselines in terms of test regret; even in the few scenarios where performance gains are more modest, the margins remain small and competitive, demonstrating reliable cross-scenario robustness without significant performance degradation.

We hope that the detailed response explicitly clarifies the contributions and benefits of the LLM-driven paradigm of EvoMIDFL.

AuthorsOpenReview ↗

Detailed Response to Reviewer ogv7 (Q2)

Q2: How does EvoMIDFL compare with conventional search under matched budgets?

We thank the reviewer for raising the question. We conducted experiments with three conventional search methods: random search, evolutionary search, and Bayesian optimization. In each baseline, a conventional search method replaces the Actor's proposal step. With the same number of validation rounds, EvoMIDFL–Full has lower mean test regret than each of these baselines at all four scales. Equal validation rounds do not establish equal computational budgets.

As reported in Table 3 of the revised manuscript, EvoMIDFL–Full achieves the lowest mean test regret across all four benchmarks:

  • Substantial Gains on Complex Matching Tasks: Relative to the strongest baseline at each scale, EvoMIDFL–Full reduces mean test regret by 7.4% on Matching-900 (from 47.862 to 44.304) and 9.2% on Matching-1600 (from 70.473 to 64.011), while also drastically lowering run-to-run variance (e.g., standard deviation drops from 1.533 to 0.364 on Matching-900 and from 5.480 to 0.333 on Matching-1600).
  • Consistent Improvements on Knapsack Tasks: On Knapsack-1000 and Knapsack-2000, EvoMIDFL–Full yields consistent reductions in mean test regret per item, from 0.266 to 0.263 (1.1%) and from 0.265 to 0.261 (1.5%), respectively, while maintaining highly competitive test RMSE.

These results demonstrate that conditioning policy proposals on training diagnostics enables EvoMIDFL to identify more effective gradient policies than black-box scalar search, yielding superior final decision quality.

Table 3: Comparison with conventional search baselines on Matching and Knapsack (in main paper). RS, ES, and BO denote random search, evolutionary search, and Bayesian optimization. Knapsack regret is per item.

MethodMatching-900 regret ↓Matching-900 RMSE ↓Matching-1600 regret ↓Matching-1600 RMSE ↓Knapsack-1000 regret/item ↓Knapsack-1000 RMSE ↓Knapsack-2000 regret/item ↓Knapsack-2000 RMSE ↓
RS51.254 ± 0.7048.736 ± 1.60472.654 ± 6.0419.620 ± 0.9470.289 ± 0.0222.260 ± 0.3590.304 ± 0.0642.439 ± 0.647
ES49.559 ± 4.8638.441 ± 2.15470.473 ± 5.4809.742 ± 0.8310.266 ± 0.0111.885 ± 0.0050.265 ± 0.0092.174 ± 0.132
BO47.862 ± 1.5339.372 ± 1.10672.429 ± 2.0407.126 ± 1.2510.270 ± 0.0231.979 ± 0.1270.268 ± 0.0043.604 ± 0.399
EvoMIDFL–Full (Ours)44.304 ± 0.3647.204 ± 0.43164.011 ± 0.3336.939 ± 0.9780.263 ± 0.0351.873 ± 0.2180.261 ± 0.0242.025 ± 0.237
Reviewer_vp3COpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes EvoMIDFL, an LLM-driven self-evolving framework for mixed-integer decision-focused learning (MIDFL). It employs an actor–critic workflow to adaptively select and combine surrogate backward-gradient mechanisms, supported by validation-guided evolution and a three-layer memory architecture. Experiments on bipartite matching and knapsack problems demonstrate improved downstream decision quality over fixed gradient methods, with additional analyses of memory transfer and policy evolution.

Critical strengths

1、The paper addresses an important challenge in decision-focused learning, where different surrogate gradient estimators may exhibit varying effectiveness across problem structures and training stages.

2、EvoMIDFL provides a structured framework integrating gradient-policy representation, candidate generation, validation-based selection, and memory-guided evolution.

3、The experiments consider different problem scales, prediction models, training-data availability, and memory configurations.

Critical weaknesses

1、The necessity and distinct contribution of LLM-driven evolution remain insufficiently established.

2、 The performance improvements over the strongest baselines are relatively modest.

3、 The manuscript does not provide a sufficiently detailed comparison of end-to-end computational costs.

The latter two concerns are particularly important: whether introducing an LLM is truly necessary, and whether the resulting performance improvements are sufficient to justify the substantial additional search costs.

Questions for authors
  1. Although EvoMIDFL is presented as an agentic self-evolving framework, its underlying optimization process closely resembles automated algorithm selection and configuration within a predefined gradient-policy search space. However, the experiments primarily compare EvoMIDFL against fixed backward-gradient methods, without considering conventional automated search strategies such as evolutionary algorithms, random search, or Bayesian optimization. Consequently, it remains unclear whether the observed improvements stem from LLM-based reasoning or simply from iterative search and validation.

The authors should explicitly justify the necessity and benefits of introducing LLMs, particularly their advantages over simpler, non-LLM alternatives.

  1. Although EvoMIDFL achieves competitive performance across several benchmarks, its improvements over the strongest fixed-backward baselines are relatively modest in many cases. For instance, the differences are small on Matching-225 and Matching-900, while EvoMIDFL does not achieve the best result on Matching-1600. Furthermore, although Appendix D.1.1 mentions five independent runs, the main tables report only average results without variability estimates or statistical significance tests. This makes it difficult to assess the reliability and practical significance of the reported improvements.

  2. EvoMIDFL requires repeated candidate evaluations, additional training, and MILP solver calls during policy evolution, potentially introducing substantial computational overhead. A comprehensive cost analysis is needed, including wall-clock time, solver evaluations, and LLM inference overhead. More importantly, whether the achieved performance improvements justify these additional costs remains insufficiently established.

  3. EvoMIDFL integrates gradient estimator selection, hyperparameter configuration, stabilization, and composition into a unified framework. However, the individual contributions of these components remain unclear. More targeted ablation studies are needed to distinguish their respective effects, particularly whether dynamic gradient composition consistently provides additional benefits beyond selecting and configuring a single estimator.

  4. The manuscript introduces extensive mathematical notation and formalization for several relatively straightforward operations within the agentic search and validation workflow, making the core methodological contributions unnecessarily difficult to follow. Additionally, several experimental figures, particularly those illustrating evolution trajectories and mechanism analyses, suffer from excessively small text and dense annotations. The presentation would benefit from substantial simplification and improved visual clarity.

Overall, the manuscript would benefit from a more focused presentation and stronger evidence establishing the necessity, effectiveness, and distinct contributions of its proposed components, particularly the LLM-driven evolution mechanism. The authors should prioritize demonstrating the methodological advantages and practical value of the framework over introducing additional architectural complexity and extensive formalization.

12 comments
AuthorsOpenReview ↗

General Response to Reviewer vp3C

We sincerely thank the reviewer for the detailed assessment and constructive feedback. We particularly appreciate the recognition of EvoMIDFL's gradient-policy representation and its evaluation across problem settings.

We understand that the reviewer's main concerns are the role of LLM-driven evolution and the breadth of our experimental analysis. To address these points, we have revised the manuscript as follows:

  • Role of LLM-driven evolution: We clarified how the LLM-driven Actor leverages training diagnostics to propose candidate policies and compared its performance with conventional search methods and fixed baselines.

  • Component ablation: We evaluated the individual contributions of estimator selection, hyperparameter tuning, stabilization, and policy composition through a staged ablation study.

  • Supplementary Results & Cost Analysis: We reported detailed decision-quality and variance metrics, and cost results, including wall-clock time, solver calls, and LLM token usage.

  • Enhanced Clarity: We relocated the formal policy-promotion rules section to Appendix and enhanced the legibility of all figures.

Please refer to the detailed responses below for further details on each revision. We believe these updates have substantially strengthened the rigor and readability of our paper, and we hope our responses fully address your questions.

AuthorsOpenReview ↗

Detailed Response to Reviewer vp3C (Q1)

Q1: Necessity and benefits of introducing LLMs.

We thank the reviewer for raising this important question. To thoroughly evaluate the contribution of LLM-driven evolution, we first conducted three new automated search experiments (AutoML methods): Random Search, Evolutionary Search, and Bayesian Optimization, and then clarified the benefits of LLM-driven methods.

1. Experiments of AutoML:

  • Settings: In AutoML experiments, we replaced the Actor with each AutoML method. Results demonstrate that under an identical budget of validation rounds, EvoMIDFL–Full achieves lower mean test regret than all three baselines across all settings (in Table 3: Comparison with conventional search baselines on Matching and Knapsack in the main paper).
  • Empirical gains. Empirically, the results confirm that LLM-driven, diagnostic-guided policies yield superior decision quality compared to non-LLM baselines under an identical evolutionary budget. As shown in Table 3, given the same number of evolution rounds, EvoMIDFL–Full achieves the lowest mean test regret across all four Matching and Knapsack benchmarks. Relative to the strongest AutoML baseline on each task, the most substantial regret reductions occur on the Matching benchmarks: 7.4% on Matching-900 (from 47.862 to 44.304) and 9.2% on Matching-1600. On Knapsack-1000 and -2000, the reductions in mean test regret per item are more modest, yet consistent at 1.1% and 1.5%, respectively.

2. LLM-driven benefits:

Based on the conventional AutoML results and previous comparison with fixed baselines, EvoMIDFL overcomes both (i) the blind black-box parameter tuning of conventional AutoML and (ii) the rigid, hand-crafted heuristics of fixed baselines that fail to adapt across problems or scales. The two key differences between EvoMIDFL and conventional AutoML search or fixed baselines are detailed below:

(1) Diagnostic-Guided Reasoning (EvoMIDFL v.s. AutoML)

  • How EvoMIDFL works. The LLM Actor agent explicitly generates proposals based on training diagnostics, the parent policy, and recent trial outcomes. For example, when validation regret plateaus alongside fluctuating gradient norms, the Actor can diagnose optimization instability and prioritize a coordinated candidate, such as switching the gradient estimator while simultaneously tightening the clipping threshold to directly resolve the underlying failure mode.
  • Conventional search. In contrast, conventional AutoML methods treat policy search as a black-box optimization problem relying solely on scalar validation scores. Specifically, Random Search samples uniformly, Evolutionary Search perturbs candidates near historically strong policies, and Bayesian Optimization fits a surrogate model exclusively on past validation scores without inspecting intermediate training dynamics. Crucially, conventional AutoML methods typically optimize scalar validation metrics via black-box sampling, lacking visibility into ongoing training dynamics to diagnose failure modes or perform targeted, structural policy refinements.
  • Benefits: The targeted-validation diagnosis in EvoMIDFL overcomes the blind black-box parameter tuning of conventional AutoML, which is supported by the empirical experiments of EvoMIDFL with AutoML under identical evolution rounds.

(2) Paradigm Shift Toward Dynamic, Training-Aware Policy Adaptation: (EvoMIDFL v.s. fixed baselines)

  • Conventional DFL: Conventional decision-focused learning (DFL) pipelines typically treat the backward gradient estimator as a static design choice fixed prior to training, requiring practitioners to manually select, tune, and validate candidate surrogate formulations through laborious trial and error.

  • New paradigm: EvoMIDFL shifts this paradigm by formulating gradient-policy selection as a dynamic, sequential decision process that continuously adapts alongside evolving training dynamics. By tightly coupling structured policy proposals with validation-gated promotion, the framework autonomously navigates combinatorial spaces of estimators, hyperparameters, and stabilization mechanisms. This closed-loop automation eliminates the need for exhaustive manual candidate enumeration and domain-specific hand-crafting, substantially reducing human engineering effort while adapting seamlessly to varying optimization stages and problem structures.

  • Balancing accuracy and stability: EvoMIDFL ensures both high accuracy and strong stability in final decision quality. Across the majority of evaluated settings, EvoMIDFL outperforms individual single-policy baselines in terms of test regret in most occasions; even in the few scenarios where performance gains are more modest, the margins remain small and competitive, demonstrating reliable cross-scenario robustness without significant performance degradation.

AuthorsOpenReview ↗

Detailed Response to Reviewer vp3C (Q1 cont.)

Comparison with conventional search baselines on Matching and Knapsack (main paper Table 3). RS, ES, and BO denote random search, evolutionary search, and Bayesian optimization. Knapsack regret is per item.

MethodMatching-900 regret ↓Matching-900 RMSE ↓Matching-1600 regret ↓Matching-1600 RMSE ↓Knapsack-1000 regret/item ↓Knapsack-1000 RMSE ↓Knapsack-2000 regret/item ↓Knapsack-2000 RMSE ↓
RS51.254 ± 0.7048.736 ± 1.60472.654 ± 6.0419.620 ± 0.9470.289 ± 0.0222.260 ± 0.3590.304 ± 0.0642.439 ± 0.647
ES49.559 ± 4.8638.441 ± 2.15470.473 ± 5.4809.742 ± 0.8310.266 ± 0.0111.885 ± 0.0050.265 ± 0.0092.174 ± 0.132
BO47.862 ± 1.5339.372 ± 1.10672.429 ± 2.0407.126 ± 1.2510.270 ± 0.0231.979 ± 0.1270.268 ± 0.0043.604 ± 0.399
EvoMIDFL–Full (Ours)44.304 ± 0.3647.204 ± 0.43164.011 ± 0.3336.939 ± 0.9780.263 ± 0.0351.873 ± 0.2180.261 ± 0.0242.025 ± 0.237

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to All Reviewers

We sincerely thank all reviewers for their careful evaluation and constructive feedback. We appreciate that Reviewers vp3C and ogv7 recognized the motivation for adaptive surrogate gradients and that Reviewer LiAv recognized EvoMIDFL as a coherent framework.

In this work, we present a novel paradigm named EvoMIDFL, a self-evolving agentic framework for mixed-integer decision-focused learning. Through an actor–critic workflow, EvoMIDFL automatically discovers and refines backward gradient policies via validation-guided diagnostic modifications, while leveraging a three-layer memory architecture to distill empirical evidence from prior trials and guide subsequent policy evolution.

In response to the reviewers' suggestions, we have made substantial efforts to revise the manuscript and expand the empirical analysis to improve clarity and coverage. The major updates are summarized below:

  • Contribution of LLM-Driven Evolution (Response to All Reviewers): LLM-driven evolution leverages semantic reasoning over ongoing training dynamics and validation feedback to autonomously synthesize, diagnose, and distill task-adaptive gradient policies without manual engineering, yielding superior decision quality. Specificaly, EvoMIDFL overcomes both the rigid, hand-crafted heuristics of fixed baselines that fail to adapt across problems or scales and the blind black-box parameter tuning of conventional AutoML.

  • Comparison with conventional search (Response to All Reviewers): We have incorporated random search, evolutionary algorithms, and Bayesian optimization as comparative baselines. Results confirm that our LLM-driven approach achieves superior final decision quality under equivalent evolutionary rounds.

  • Supplementary results & cost analysis (Response to Reviewers vp3C and LiAv): We reported per-seed accuracy results and a cost analysis that includes wall-clock time, solver calls, and representative LLM token usage for Matching and Knapsack benchmarks to provide a comprehensive numerical analysis.

  • New ablation & benchmark results (Response to Reviewers vp3C and LiAv): We conducted four new experiments: (i) a new submodule ablation of EvoMIDFL, (ii) a comparison across validation sets, (iii) a reverse memory transfer analysis (Knapsack to Matching), and (iv) an evaluation on an unseen MILP family (Newsvendor).

  • Enhanced presentation & visualization (Reviewer vp3C): We increased the font sizes across all figures and captions for legibility and relocated the policy-promotion rules Section to the appendix to streamline the main text.

All revisions in the updated manuscript are highlighted in blue. We believe these revisions have significantly improved the quality of our work. Detailed point-by-point responses to each reviewer are provided below.

AuthorsOpenReview ↗

Summary of Response and Discussion During the Rebuttal Phase

Dear Chairs,

TL;DR: Our paper's average score increased from 2.0 (2 / 2 / 2) to 2.67 (3 / 3 / 2) following a constructive discussion period. Two engaged reviewers (R-ogv7 and R-LiAv) raised their scores after we addressed their concerns with new experiments and clarifications.

In this work, we proposed EvoMIDFL, a novel self-evolving agentic paradigm for mixed-integer decision-focused learning aiming to improve final decision quality. Driven by an actor–critic workflow, EvoMIDFL automatically discovers and refines backward gradient policies through validation-guided diagnostic modifications, supported by a three-layer memory architecture that distills insights from prior trials. By eliminating manual surrogate selection, EvoMIDFL achieves strong decision performance across diverse scenarios and problem scales.

The table below summarizes the review process. All score updates were made by the reviewers following our rebuttal and follow-up discussions:

Reviewer IDInitial ScorePost-Discussion ScoreConfidence
R-LiAv233
R-ogv7232
R-vp3C224

Summary of the Rebuttal Phase

Our rebuttal and subsequent discussions were highly constructive, leading to substantial improvements in the revised manuscript. We are grateful to the reviewers for their diligent and fair evaluation, with all interactions conducted in full accordance with academic integrity. Below is a brief summary of how we addressed each reviewer's key concerns and their responses during the discussion phase:

  • Reviewer LiAv (2 →\rightarrow 3): Reviewer LiAv's main questions focused on experimental breadth and robustness. In response, we reported per-seed statistics and computational cost analyses, and added supplementary experiments on validation-set sensitivity, reverse cross-task memory transfer, and an unseen MILP problem family (Newsvendor). These new experiments substantially strengthened our work's empirical rigor. Following these additions, Reviewer LiAv commented "Thanks for the careful rebuttal and additional experiments", and raised their score from 2 to 3.

  • Reviewer ogv7 (2 →\rightarrow 3): Reviewer ogv7's primary concern centered on the specific contribution of LLM-driven evolution. Following the reviewer's suggestions, we clarified the Actor's role in diagnostic-guided reasoning and added two new experiments: comparisons against conventional search methods and an ablation without failure diagnostics. These results confirm consistent regret reductions by EvoMIDFL across all benchmarks. In response, Reviewer ogv7 appreciated the comparisons, replied "Good!", and raised their score from 2 to 3.

  • Reviewer vp3C (2): Reviewer vp3C raised questions regarding the contribution of LLM-driven evolution and the depth of experimental analysis. We clarified EvoMIDFL's core contribution as a paradigm shift toward dynamic, diagnostic-guided reasoning for targeted policy refinement rather than manual heuristic design. Additionally, we added submodule ablations, per-seed statistics, computational cost, and LLM usage analyses. Although Reviewer vp3C did not respond during the discussion period, we believe our revisions thoroughly address their concerns.

We are deeply grateful to the reviewers for their time, expertise, and constructive engagement, which have significantly strengthened our paper. We trust that this summary, together with the full rebuttal record, provides a clear overview of the positive peer-review discussion our submission has undergone.

Thank you for your time and dedicated service to the community. We look forward to your final decision.

Sincerely,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.