Pith. sign in

REVIEW 4 major objections 8 minor 2 cited by

DHEvo: Data-Algorithm Based Heuristic Evolution for Generalizable MILP Solving

T0 review · 4 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read DHEvo claims that iteratively selecting the instances on which current diving heuristics perform best, then evolving new heuristics on those instances, yields diving rules that generalize across a MILP problem class and beat both…

desk verdict Co-evolving instances with algorithms is a genuine and useful idea; the evidence is solid on held-out sets, but the 'representativeness' hypothesis is unproven and the selection bias deserves a closer look. read the letter →

arxiv 2507.15615 v1 pith:X5WQSPPL submitted 2025-07-21 cs.NE

Zhihao Zhang , Siyuan Li , Chenxi Li , Feifan Liu , Mengjing Chen , Kai Li , Tao Zhong , Bo An
show 1 more author
Peng Liu
This is my paper
classification cs.NE MSC 90C1190C5968W50
keywords mixedintegerprogrammingprimalheuristicsdivingevolutionarycomputationlargelanguagemodelsdata-algorithmco-evolutiongeneralizationmulti-agentsystem
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that primal heuristics for mixed-integer linear programming (MILP) can be made to generalize across a whole problem class if the heuristic and the training instances are evolved together, instead of evolving the heuristic on a fixed, small set of instances. Its framework, DHEvo, repeatedly keeps the (instance, heuristic) pairs with the best relative primal gap, treats those instances as structurally representative, and uses them to seed the next generation of LLM-generated diving heuristics. On four combinatorial benchmarks the paper reports lower average primal gap and lower variance than hand-crafted diving rules, a learned baseline, and prior LLM-evolution methods, and when the evolved rules are plugged into a full solver they also reduce solving time and primal-dual integral. If this holds, automatically adapting primal heuristics to an instance distribution without expert tuning becomes a realistic route to faster MILP solving.

What carries the argument

The load-bearing object is the data-code pair: an instance from the problem class plus a diving heuristic coded as a Python scoring function. A diving heuristic is just a rule that reads 13 features of the LP-relaxation solution of each fractional variable (fractional value, objective coefficient, pseudocost, lock counts, and similar) and returns a score and a rounding direction; the central machinery is the evolution loop around those pairs. Each generation evaluates the current heuristic population on the current instance set, keeps the pairs with the best relative primal gap, and feeds the surviving heuristic code back into a four-role LLM agent team (designer, coder, reviewer, judge) that performs mutation and crossover through prompts. A temperature-controlled retention step and a final average-fitness selection over the retained instances close the loop.

What would settle it

Construct a synthetic problem class with two equally sized clusters of instances that look similar in features but differ sharply in difficulty (one with tight LP relaxations, one with loose). Run DHEvo with access to both clusters and test the final heuristic on held-out instances of each. If the heuristic's relative primal gap is much better on the easy cluster and no better than a random diving rule on the hard cluster, the fitness-based selection is choosing easy instances rather than representative ones, and the generalization claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that selection pressure on instances is as important as selection pressure on algorithms. DHEvo starts by sampling instances from a problem-class distribution, generates an initial population of diving heuristics with an LLM-based multi-agent system, evaluates each heuristic on each sampled instance by relative primal gap, and keeps the top instance-heuristic pairs. In later generations those surviving instances, not a fresh random draw, are the training set for the next round of crossover and mutation, so the data and the code co-adapt. The reported result is consistent improvement: on the independent-set benchmark the average primal gap improves 56.04% over the best hand-crafted diving heuristic, on the set-cover benchmark the best LLM-evolution baseline is beaten by 61.8%, and DHEvo shows the lowest performance variance on all four datasets. In solver-integration experiments the evolved diving rules reduce solving time and primal-dual integral relative to default and tuned solver settings.

Load-bearing premise

The load-bearing premise is that an instance on which the current heuristic scores well is structurally representative of the whole problem class, so evolving on those instances improves performance on unseen instances too; if that premise fails, selection just biases evolution toward easy instances.

Editorial extensions

If this is right

  • If the co-evolution claim is right, diving heuristics can be customized to a problem class automatically, without a human expert hand-tuning scoring rules for each class.
  • Generalization within a class should improve: the paper measures both lower average relative primal gap and lower variance on all four benchmarks, with variance reductions as large as 46.9% on set-cover relative to the best LLM baseline.
  • Integrating the evolved dives into a full solver should shorten time-to-solution and reduce primal-dual integral, not just improve the isolated diving metric.
  • The ablation suggests the gain is not merely from the LLM generator: adding the co-evolution loop to a baseline LLM-evolution pipeline also reduces variance, by nearly 30% on set-cover, so instance selection is the main driver.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural extension the paper does not develop: fitness-based instance selection is effectively a self-paced curriculum, so the same data-algorithm loop could be applied to other solver components with measurable per-instance effects, such as branching rules, cut selection, or node-selection policies.
  • Because representativeness is defined by the current heuristic's fitness, the loop carries an implicit bias toward instances with tight LP relaxations; a testable alternative is to define representativeness from instance features alone and compare which selection rule generalizes better on held-out hard instances.
  • A practical variation would be to cluster instances within a problem class and run one co-evolution per cluster, which may handle multi-modal instance distributions better than a single population and could be evaluated on the same four benchmarks.
  • The multi-agent debate component and the co-evolution component are separable; rerunning the experiments with a cheaper single-agent generator would indicate how much of the reported gain comes from the agent team versus the data selection mechanism.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper proposes DHEvo, a framework for automatically generating generalizable diving heuristics for mixed-integer linear programming (MILP). The method co-evolves a population of MILP instances and heuristic scoring functions: starting from a randomly sampled instance set and an LLM-generated heuristic population, it evaluates each heuristic per instance, selects the top-k instance-heuristic pairs by fitness (relative primal gap), and iteratively re-evolves heuristics on those selected pairs. The final heuristic set is chosen by averaging performance over the selected instances. The authors evaluate on four synthetic benchmarks (cauctions, setcover, facilities, indset) and three real-world datasets, comparing against human-designed SCIP heuristics, L2DIVE, and LLM-based evolutionary methods (FunSearch, EoH, HillClimb, LLM4Solver). They report improvements in average primal gap, performance variance, solving time, and primal-dual integral, and argue that their data-algorithm co-evolution mechanism improves generalization within a problem class.

Significance. If the central claim holds, the paper offers a meaningful advance: an automated, LLM-driven approach that actively selects training instances during heuristic evolution, rather than treating a fixed sample as representative. The evaluation is commendable for using held-out test sets, multiple benchmarks, and several strong baselines. The method is not circular in the narrow sense of tuning on the test set, and the appendix provides useful detail on prompts, features, and generated heuristics. However, the paper's load-bearing assumption—that high fitness indicates structural representativeness—is not established, and the empirical analysis does not yet rule out the alternative that the selection procedure simply picks easy instances and thereby biases the heuristics toward them. The statistical support for the variance claims is also incomplete. With additional evidence addressing these points, the paper could become a solid contribution.

major comments (4)
  1. [Section 3.1, Algorithm 1 (lines 6-11, 18, 21)] The selection of 'representative' instances is operationalized by ranking instances according to Perf(I_i, h*_i), i.e., the performance of the best heuristic found for that instance. An instance is selected precisely when some current heuristic already solves it well, which makes the selection a difficulty filter rather than a representativeness filter. The paper's Insight 1 does not bridge this gap: the argument that a small integrality gap implies that heuristics trained on that instance generalize well is asserted rather than proven, and a small integrality gap does not imply that the heuristic's performance on that instance predicts performance on other instances. Because the final selection (line 21) averages only over the already selected instances, the reported lower variance on held-out sets could be an artifact of specializing to easy instances. The authors should provide evidence that selected instances are not systematically easier than unselected ones, for example by comparing the difficulty distribution of selected versus rejected instances, or by reporting the performance of the final heuristics on a separate set of hard instances from the same problem class.
  2. [Section 4.2, Table 1] The variance comparison is not statistically grounded. The table header says 'standard error' but the parenthetical values appear to be standard deviations of the primal gap across test instances; the text refers to 'performance variance' and 'standard error' interchangeably. No confidence intervals, paired tests, or equality-of-variance tests are reported, and with only three random seeds the variability of the evolutionary process itself is uncharacterized. The claim of 'lowest variance across all four datasets' is therefore not quantitatively substantiated. The authors should report per-seed results, confidence intervals, and an appropriate statistical test (e.g., Levene's test or a paired bootstrap) for both means and variances.
  3. [Section 4.2, Table 1, L2DIVE row] The comparison with L2DIVE is not controlled. Appendix B states that L2DIVE's performance is taken from the original paper because the code is not open-source, rather than being run in the same environment with the same features, time limits, and solver settings. Consequently, the claim that DHEvo outperforms a 'learning-based GNN method' is not supported by the experiments reported here. The authors should either run L2DIVE under identical conditions or remove it from the comparison table.
  4. [Section 3.1, Insights 1 and 2] The two 'Insights' are presented as established facts, but they are unproven hypotheses; the Introduction even states 'we assume.' Since these assumptions drive the entire selection mechanism, they should either be proven, empirically tested, or clearly labeled as assumptions with supporting evidence. For example, the paper could test Insight 1 by comparing the generalization of heuristics evolved on instances with high versus low integrality gaps, or by measuring whether the selected instances have features that are actually representative of the wider problem-class distribution.
minor comments (8)
  1. [Table 3] The header 'Cautions' should be 'Cauctions'.
  2. [Algorithm 1, line 18] The word 'smale' should be 'sample' in 'smale top-k pairs P* ← Smaple(...)'.
  3. [Figure 1] The label 'Temperture' should be 'Temperature'.
  4. [Section 4.4] The sentence 'the variance of the evolved heuristics decreases by nearly 30% on the setcover dataset when the co-evolution mechanism is removed' contradicts both the preceding sentence (which says variance increases by about 10% when co-evolution is excluded) and the numbers in Table 3 (DHEvo std 13.42, DHEvo_OFF std 13.99; EoH std 28.89, EoH_DH std 17.48). This needs correction and clarification.
  5. [Section 2.2] The primal-dual gap definition restricts to '0 < z, z* < ∞', which excludes zero or negative objective values that are common in MILP; please clarify how the gap is computed in those cases.
  6. [Table 2 caption] The caption says 'SCIPP' and should say 'SCIP'.
  7. [Figure 2] The explanation that Setcover and Facilities appear as single points in the t-SNE visualization requires more detail, since the datasets include both easy and hard instances with varying sizes.
  8. [Appendix B] The dataset names 'MILPLIB' and 'MIPLIB' are used inconsistently; please unify the terminology.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the generalization claim is tested on held-out instances, so the fitness-based instance selection is not statistically forced.

full rationale

The paper's central claim is that DHEvo evolves diving heuristics that generalize across a problem class. The selection mechanism in Algorithm 1 ranks instance-heuristic pairs by Perf(I_i, h*_i) and keeps the top-k pairs, but the discovered heuristics are then evaluated on separately generated test instances (100 per dataset) that were not used during evolution. This means the measured generalization is an extrapolation, not a re-reporting of the fitness values used for selection. The two 'Insights' in Section 3.1 are explicitly presented as hypotheses ('we hypothesize that those with higher fitness scores ... will likely exhibit greater structural representativeness'), and they are not derived from the paper's own equations by construction. Even though Insight 1 is not rigorously proven and may be a correctness risk, an unproven assumption is not the same as circularity. The paper contains no load-bearing self-citations: the cited LLM-evolution baselines are external works with no author overlap, and the method's components (MA-Evolution System, co-evolution, prompts) are evaluated through ablations and held-out benchmarks. Therefore, no prediction reduces to its input by definition or by fitted parameters, and the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method rests on unproven theoretical insights about instance representativeness and on several unspecified hyperparameters. The evolved heuristics contain numerical coefficients that are fitted during search, but these are outputs rather than ad hoc inputs to the framework.

free parameters (3)
  • Population size m and number of training instances n = Not specified
    Algorithm 1 requires these as inputs, but the paper does not report the values used in experiments, which affect the search and generalization.
  • Top-k and temperature T in selection = Not specified
    The retention probability formula includes temperature T, and top-k determines how many data-code pairs survive; no values are given.
  • Coefficients in evolved scoring functions = Evolved by LLM search on training instances
    The final heuristics (Appendix E) contain numeric weights (e.g., 80, 50, 90, 25) that are fitted through the evolutionary process, not derived from first principles.
assumptions (4)
  • ad hoc to paper Insight 1: Instances with high fitness (low primal gap) are structurally representative and lead to lower variance on similar instances.
    Stated in Section 3.1 without proof; it is the core justification for the instance selection step.
  • ad hoc to paper Insight 2: Instances with more regular feasible regions yield heuristics that generalize better.
    Also stated in Section 3.1 without proof; used to justify selection of 'representative' instances.
  • domain assumption The relative primal gap is an appropriate fitness measure for diving heuristics.
    The paper uses Perf(h,I) = relative primal gap, assuming it correlates with heuristic quality; standard in the field but still an assumption.
  • domain assumption LLM-generated code can implement valid diving rules.
    The MA-Evolution System relies on the LLM producing syntactically correct and semantically reasonable scoring functions; the paper checks this by compiling and running, but it is a practical assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DHEvo: Data-Algorithm Based Heuristic Evolution for Generalizable MILP Solving." pith.science (2026). https://pith.science/paper/X5WQSPPL

@misc{pith2026250715615,
  author       = {Pith},
  title        = {Pith review of: DHEvo: Data-Algorithm Based Heuristic Evolution for Generalizable MILP Solving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X5WQSPPL}},
  note         = {Machine review of arXiv:2507.15615}
}
read the original abstract

Primal heuristics play a critical role in improving the efficiency of mixed integer programming (MILP) solvers. As large language models (LLMs) have demonstrated superior code generation abilities, recent MILP works are devoted to leveraging the evolutionary computation approaches with LLMs to generate effective primal heuristics. Although the generated heuristics have achieved better solving performance than the hand-crafted ones with little adaptability, the advantage of current LLM-based methods is limited to few MILP instances in one problem class, as they fail to capture the instance characteristics in the problem class (the MILP instances generated from the same mathematical model are defined as a problem class). Since MILP instances often differ significantly in structure and feature distribution, the neglect of their characteristics in the evolution process results in poor generalization within the same problem class. To overcome this challenge, we propose a data-algorithm co-evolution framework (DHEvo) that iteratively selects representative instances and evolves corresponding heuristics. With the initial instance distribution, we develop an LLM-based multi-agent system to generate data-code pairs simultaneously. These data-code pairs are iteratively refined based on their fitness scores, leading to the identification of the most effective heuristic over the entire problem class. Extensive experiments across diverse MILP benchmarks demonstrate that our approach significantly outperforms both human-designed heuristics and existing LLM-based methods.

Figures

Figures reproduced from arXiv: 2507.15615 by the authors.

Figure 1
Figure 1. Illustration of data-algorithm co-evolution framework (DHEvo). [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The visualization of instance features via t-SNE is presented as follows: Panel a represents [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Illustration of MA-Evolution System. 1, it begins by selecting a subset of high-quality instance-heuristic pairs (I ∗ , H∗ ), where each heuristic in H∗ demonstrates strong performance on its corresponding instance in I ∗ . These pairs are identified by evaluating all candidate heuristics against a sampled MILP dataset and ranking them based on a independent fitness score. These instances with higher scores indicate… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The prompts of contextual information of MILP. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: The prompts in MA-Evolution System. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. AHD Agent: Agentic Reinforcement Learning for Automatic Heuristic Design

    cs.AI 2026-05 unverdicted novelty 7.0 of 10

    AHD Agent trains a 4B-parameter LLM via agentic RL to actively use tools for automatic heuristic design, matching or exceeding larger baselines across eight domains with fewer evaluations.

  2. LLM4Branch: Large Language Model for Discovering Efficient Branching Policies of Integer Programs

    cs.AI 2026-05 unverdicted novelty 6.0 of 10

    LLM4Branch discovers branching policies for MILP solvers as LLM-generated executable programs whose parameters are tuned via zeroth-order optimization on solver performance.

Reference graph

Works this paper leans on

47 extracted references · 32 canonical work pages · cited by 2 Pith papers

  1. [1]

    A tsp-based milp model for medium-term planning of single-stage continuous multiproduct plants.Industrial & Engineering Chemistry Research, 47(20):7733–7743, 2008

    Songsong Liu, Jose M Pinto, and Lazaros G Papageorgiou. A tsp-based milp model for medium-term planning of single-stage continuous multiproduct plants.Industrial & Engineering Chemistry Research, 47(20):7733–7743, 2008

    work page 2008

  2. [2]

    Biodiesel supply chain optimization modeled with geographical information system (gis) and mixed-integer linear programming (milp) for the northern great plains region

    Hyunju Jeong, Heidi L Sieverding, and James J Stone. Biodiesel supply chain optimization modeled with geographical information system (gis) and mixed-integer linear programming (milp) for the northern great plains region. BioEnergy research, 12:229–240, 2019

    work page 2019

  3. [3]

    An milp model for optimization of a small-scale lng supply chain along a coastline

    Raine Jokinen, Frank Pettersson, and Henrik Saxén. An milp model for optimization of a small-scale lng supply chain along a coastline. Applied energy, 138:423–431, 2015

    work page 2015

  4. [4]

    Accelerating an fpga-based sat solver by software and hardware co-design

    Kefan Ma, Liquan Xiao, Jianmin Zhang, and Tiejun Li. Accelerating an fpga-based sat solver by software and hardware co-design. Chinese Journal of Electronics, 28(5):953–961, 2019

    work page 2019

  5. [5]

    Constraint improvements for milp-based hardware synthesis

    Lou Hafer. Constraint improvements for milp-based hardware synthesis. In Proceedings of the 28th ACM/IEEE Design Automation Conference, pages 14–19, 1991

    work page 1991

  6. [6]

    Integrated production and outbound distribution scheduling: review and extensions

    Zhi-Long Chen. Integrated production and outbound distribution scheduling: review and extensions. Operations research, 58(1):130–148, 2010

    work page 2010

  7. [7]

    An milp for scheduling problems in an fms with one vehicle

    Anthony Caumond, Philippe Lacomme, Aziz Moukrim, and Nikolay Tchernev. An milp for scheduling problems in an fms with one vehicle. European Journal of Operational Research, 199(3):706–722, 2009

    work page 2009

  8. [8]

    Optimization of the power output scheduling of a renewables-based hybrid power station using milp approach: The case of tilos island

    Francesco Superchi, Nathan Giovannini, Antonis Moustakis, George Pechlivanoglou, and Alessandro Bianchini. Optimization of the power output scheduling of a renewables-based hybrid power station using milp approach: The case of tilos island. Renewable Energy, 220:119685, 2024

    work page 2024

Show all 47 references
  1. [9]

    A practical mixed integer linear programming based approach for unit commitment

    Gary W Chang, YD Tsai, CY Lai, and JS Chung. A practical mixed integer linear programming based approach for unit commitment. In IEEE Power Engineering Society General Meeting., pages 221–225. IEEE, 2004

    2004

  2. [10]

    Optimal sizing and energy management of a microgrid: A joint milp approach for minimization of energy cost and carbon emission

    Fadi Agha Kassab, Berk Celik, Fabrice Locment, Manuela Sechilariu, Sheroze Liaquat, and Timothy M Hansen. Optimal sizing and energy management of a microgrid: A joint milp approach for minimization of energy cost and carbon emission. Renewable Energy, 224:120186, 2024

    2024

  3. [11]

    Efficient expansion planning of modern multi-energy distribution networks with electric vehicle charging stations: A stochastic milp model

    Peyman Zare, Abdolmajid Dejamkhooy, and Iraj Faraji Davoudkhani. Efficient expansion planning of modern multi-energy distribution networks with electric vehicle charging stations: A stochastic milp model. Sustainable Energy, Grids and Networks, 38:101225, 2024

    2024

  4. [12]

    Worst-case analysis of two travelling salesman heuristics

    Hoon Liong Ong and JB Moore. Worst-case analysis of two travelling salesman heuristics. Operations Research Letters, 2(6):273–277, 1984

    1984

  5. [13]

    Pivot and shift—a mixed integer programming heuristic

    Egon Balas, Stefan Schmieta, and Christopher Wallace. Pivot and shift—a mixed integer programming heuristic. Discrete Optimization, 1(1):3–12, 2004. 10

    2004

  6. [14]

    Primal heuristics for mixed integer programs

    Timo Berthold. Primal heuristics for mixed integer programs. PhD thesis, Zuse Institute Berlin, 2006

    2006

  7. [15]

    Zi round, a mip rounding heuristic

    Chris Wallace. Zi round, a mip rounding heuristic. Journal of Heuristics, 16:715–722, 2010

    2010

  8. [16]

    Conflict-driven heuristics for mixed integer programming

    Jakob Witzig and Ambros Gleixner. Conflict-driven heuristics for mixed integer programming. INFORMS Journal on Computing, 33(2):706–720, 2021

    2021

  9. [17]

    The scip optimization suite 5.0

    Ambros Gleixner, Leon Eifler, Tristan Gally, Gerald Gamrath, Patrick Gemander, Robert Lion Gottwald, Gregor Hendel, Christopher Hojny, Thorsten Koch, Matthias Miltenberger, et al. The scip optimization suite 5.0. Optimization Online, 2017

    2017

  10. [18]

    Gurobi Optimizer Reference Manual, 2023

    Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2023

    2023

  11. [19]

    Large language models as optimizers

    Chengrun Yang, Xuezhi Wang, Yifeng Lu, Hanxiao Liu, Quoc V Le, Denny Zhou, and Xinyun Chen. Large language models as optimizers. arXiv preprint arXiv:2309.03409, 2023

    2023 arXiv

  12. [20]

    Language model crossover: Variation through few-shot prompting

    Elliot Meyerson, Mark J Nelson, Herbie Bradley, Adam Gaier, Arash Moradi, Amy K Hoover, and Joel Lehman. Language model crossover: Variation through few-shot prompting. ACM Transactions on Evolutionary Learning, 4(4):1–40, 2024

    2024

  13. [21]

    Evoprompting: Language models for code-level neural architecture search

    Angelica Chen, David Dohan, and David So. Evoprompting: Language models for code-level neural architecture search. Advances in neural information processing systems, 36:7787–7817, 2023

    2023

  14. [22]

    Mathematical discoveries from program search with large language models

    Bernardino Romera-Paredes, Mohammadamin Barekatain, Alexander Novikov, Matej Balog, M Pawan Kumar, Emilien Dupont, Francisco JR Ruiz, Jordan S Ellenberg, Pengming Wang, Omar Fawzi, et al. Mathematical discoveries from program search with large language models. Nature, 625(7995...

    2024

  15. [23]

    A systematic survey on large language models for algorithm design

    Fei Liu, Yiming Yao, Ping Guo, Zhiyuan Yang, Zhe Zhao, Xi Lin, Xialiang Tong, Mingxuan Yuan, Zhichao Lu, Zhenkun Wang, et al. A systematic survey on large language models for algorithm design. arXiv preprint arXiv:2410.14716, 2024

  16. [24]

    Evolution of heuristics: Towards efficient automatic algorithm design using large language model

    Fei Liu, Xialiang Tong, Mingxuan Yuan, Xi Lin, Fu Luo, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Evolution of heuristics: Towards efficient automatic algorithm design using large language model. arXiv preprint arXiv:2401.02051, 2024

    2024 arXiv

  17. [25]

    Llm4solver: Large language model for efficient algorithm design of combinatorial optimization solver

    Yuyan Zhou, Jie Wang, Yufei Kuang, Xijun Li, Weilin Luo, Jianye HAO, and Feng Wu. Llm4solver: Large language model for efficient algorithm design of combinatorial optimization solver. https://openreview.net/pdf?id=XTxdDEFR6D, 2024

    2024

  18. [26]

    Self-paced cur- riculum learning

    Lu Jiang, Deyu Meng, Qian Zhao, Shiguang Shan, and Alexander Hauptmann. Self-paced cur- riculum learning. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 29, 2015

    2015

  19. [27]

    Meta-learning for semi-supervised few-shot classification

    Mengye Ren, Eleni Triantafillou, Sachin Ravi, Jake Snell, Kevin Swersky, Joshua B Tenen- baum, Hugo Larochelle, and Richard S Zemel. Meta-learning for semi-supervised few-shot classification. arXiv preprint arXiv:1803.00676, 2018

    2018 arXiv

  20. [28]

    Learning to sample hard instances for graph algorithms

    Ryoma Sato, Makoto Yamada, and Hisashi Kashima. Learning to sample hard instances for graph algorithms. In Asian Conference on Machine Learning, pages 503–518. PMLR, 2019

    2019

  21. [29]

    Curriculum learning

    Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In Proceedings of annual international conference on machine learning, pages 41–48, 2009

    2009

  22. [30]

    An automatic method for solving discrete programming problems

    Ailsa H Land and Alison G Doig. An automatic method for solving discrete programming problems. In 50 Years of Integer Programming 1958-2008: From the Early Years to the State-of-the-Art, pages 105–132. Springer, 2009

    1958

  23. [31]

    Handbook of evolutionary computa- tion

    Thomas Bäck, David B Fogel, and Zbigniew Michalewicz. Handbook of evolutionary computa- tion. Release, 97(1):B1, 1997

    1997

  24. [32]

    Evolutionary learning: Advances in theories and algorithms

    Zhi-Hua Zhou, Yang Yu, and Chao Qian. Evolutionary learning: Advances in theories and algorithms. Springer, 2019. 11

    2019

  25. [33]

    From evolutionary computation to the evolution of things

    Agoston E Eiben and Jim Smith. From evolutionary computation to the evolution of things. Nature, 521(7553):476–482, 2015

    2015

  26. [34]

    A comprehensive overview of large language models

    Humza Naveed, Asad Ullah Khan, Shi Qiu, Muhammad Saqib, Saeed Anwar, Muhammad Usman, Naveed Akhtar, Nick Barnes, and Ajmal Mian. A comprehensive overview of large language models. arXiv preprint arXiv:2307.06435, 2023

    2023 arXiv

  27. [36]

    Evolutionary com- putation in the era of large language model: Survey and roadmap

    Xingyu Wu, Sheng-hao Wu, Jibin Wu, Liang Feng, and Kay Chen Tan. Evolutionary com- putation in the era of large language model: Survey and roadmap. IEEE Transactions on Evolutionary Computation, 2024

    2024

  28. [37]

    Towards understanding generalization of deep learning: Perspec- tive of loss landscapes

    Lei Wu, Zhanxing Zhu, et al. Towards understanding generalization of deep learning: Perspec- tive of loss landscapes. arXiv preprint arXiv:1706.10239, 2017

    2017 arXiv

  29. [38]

    How does loss function affect generalization performance of deep learning? application to human age estimation

    Ali Akbari, Muhammad Awais, Manijeh Bashar, and Josef Kittler. How does loss function affect generalization performance of deep learning? application to human age estimation. In International Conference on Machine Learning, pages 141–151. PMLR, 2021

    2021

  30. [39]

    Fantastic generalization measures and where to find them

    Yiding Jiang, Behnam Neyshabur, Hossein Mobahi, Dilip Krishnan, and Samy Bengio. Fantastic generalization measures and where to find them. arXiv preprint arXiv:1912.02178, 2019

    1912 arXiv

  31. [40]

    Encouraging divergent thinking in large language models through multi- agent debate

    Tian Liang, Zhiwei He, Wenxiang Jiao, Xing Wang, Yan Wang, Rui Wang, Yujiu Yang, Shuming Shi, and Zhaopeng Tu. Encouraging divergent thinking in large language models through multi- agent debate. arXiv preprint arXiv:2305.19118, 2023

    2023 arXiv

  32. [41]

    Chateval: Towards better llm-based evaluators through multi-agent debate

    Chi-Min Chan, Weize Chen, Yusheng Su, Jianxuan Yu, Wei Xue, Shanghang Zhang, Jie Fu, and Zhiyuan Liu. Chateval: Towards better llm-based evaluators through multi-agent debate. arXiv preprint arXiv:2308.07201, 2023

    2023 arXiv

  33. [42]

    Aflow: Automating agentic workflow generation

    Jiayi Zhang, Jinyu Xiang, Zhaoyang Yu, Fengwei Teng, Xionghui Chen, Jiaqi Chen, Mingchen Zhuge, Xin Cheng, Sirui Hong, Jinlin Wang, et al. Aflow: Automating agentic workflow generation. arXiv preprint arXiv:2410.10762, 2024

    2024 arXiv

  34. [43]

    Improving multi-agent debate with sparse communication topology

    Yunxuan Li, Yibing Du, Jiageng Zhang, Le Hou, Peter Grabowski, Yeqing Li, and Eugene Ie. Improving multi-agent debate with sparse communication topology. arXiv preprint arXiv:2406.11776, 2024

    2024 arXiv

  35. [44]

    Constraint integer programming

    Tobias Achterberg. Constraint integer programming. PhD thesis, technical university of berlin, 2007

    2007

  36. [45]

    Learning to dive in branch and bound

    Max Paulus and Andreas Krause. Learning to dive in branch and bound. Advances in Neural Information Processing Systems, 36:34260–34277, 2023

    2023

  37. [46]

    Understanding the importance of evolutionary search in automated heuristic design with large language models

    Rui Zhang, Fei Liu, Xi Lin, Zhenkun Wang, Zhichao Lu, and Qingfu Zhang. Understanding the importance of evolutionary search in automated heuristic design with large language models. In International Conference on Parallel Problem Solving from Nature, pages 185–202. Springer, 2024

    2024

  38. [47]

    The underlying similarity of diversity measures used in evolutionary computation

    Mark Wineberg and Franz Oppacher. The underlying similarity of diversity measures used in evolutionary computation. In Genetic and Evolutionary Computation, pages 1493–1504. Springer, 2003

    2003

  39. [48]

    and "mayroundup

    Adel Nikfarjam, Jakob Bossek, Aneta Neumann, and Frank Neumann. Entropy-based evolution- ary diversity optimisation for the traveling salesperson problem. In Proceedings of the Genetic and Evolutionary Computation Conference, pages 600–608, 2021. 12 A Diving Heuristics Diving ...

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.