Pith. sign in

REVIEW 4 major objections 5 minor 78 references

Neural Genetic Search in Discrete Spaces

T0 review · 4 major / 5 minor · reviewed 2026-08-08 · deepseek-v4-flash

Pith's one-line read Neural Genetic Search claims that a single token-restriction crossover turns any sequential generative model into an effective test-time search algorithm, reporting gains on routing, red-teaming, and molecular design.

desk verdict NGS is a simple, well-tested search wrapper for sequential generative models, but the missing mutation-only control leaves the headline crossover's contribution unproven. read the letter →

arxiv 2502.10433 v2 pith:ETCJDQD2 submitted 2025-02-09 cs.NE cs.LG

classification cs.NEcs.LG
keywords neuralgeneticsearchtest-timetoken-restrictioncrossoverparent-conditionedgenerationsequentialgenerativemodelsroutingproblemsred-teaminglanguagemoleculardesign
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

This paper tries to establish that genetic search and deep generative decoding can be the same operation: crossover is simply parent-conditioned generation. The proposed Neural Genetic Search (NGS) keeps a population of candidate sequences, selects two parents, and at each generation step reweights the pretrained policy's next-token distribution so that only tokens appearing in either parent are allowed. A stochastic mutation removes the restriction, and the population is iteratively replaced using rank-based sampling. Across routing problems, adversarial prompt generation for language models, and de novo molecular design, the authors report that NGS finds higher-reward solutions under the same evaluation budget than sampling, beam search, Monte Carlo tree search, and ant colony optimization built on the same policy. If that claim holds, NGS is a drop-in, problem-agnostic way to add search to any model that generates discrete outputs one token at a time.

What carries the argument

The load-bearing object is the token-restriction crossover of Eq. (2), which reweights the pretrained policy's next-token distribution so that only tokens present in either selected parent survive, together with the stochastic mutation of Eq. (3) that removes the mask when constraints require it or with a small probability otherwise. This operator carries the argument because it turns an arbitrary factorized generative policy into a genetic algorithm's variation step without any problem-specific design, and it is the only component that changes across domains; the rest of NGS is a standard population loop with rank-based prioritized sampling for selection and replacement.

What would settle it

A decisive test is to run NGS against best-of-N sampling on a routing instance while artificially degrading the policy, for example by training it only on low-quality tours; if NGS's gap advantage disappears or turns negative, the central claim fails for policies without good solutions in their support, which is the case the paper's limitation concedes.

Watch

Extended reading notes

Core claim

The paper's central claim is that the evolutionary crossover operator can be defined entirely inside the generative model. Given two parent sequences $s_1$ and $s_2$, a child is sampled from the restricted next-token distribution $$p_{\mathrm{cross}}(s_t \mid s_{<t}) \propto \mathbb{1}(s_t \in V_{s_1,s_2}) \, p_{\$\theta$^*}(s_t \mid s_{<t}),$$ where $V_{s_1,s_2}$ is the union of tokens in the two parents, and mutation is the same distribution with the mask removed. The paper then reports that this operator, wrapped in a population loop with rank-based selection, produces significantly smaller optimality gaps than sampling, beam search, MCTS, and ACO on TSP, CVRP, PCTSP, and OP; higher transfer toxicity for red-teaming attacks; and higher Top-10 scores than specialized genetic algorithms on 8 of 10 molecular-design tasks.

Load-bearing premise

The load-bearing premise is that the pretrained policy, even after its vocabulary is restricted to the two parents' tokens, still assigns enough probability mass to valid, high-quality continuations; the paper states that if the pretrained model's distribution does not encompass high-quality solutions, NGS's potential gains will be limited.

Editorial extensions

If this is right

  • NGS becomes a general decoding strategy: any trained autoregressive or otherwise factorized discrete generator can be searched at test time by adding a population and a token-masking rule, with no task-specific operator design.
  • Under the same pretrained policy and comparable budget, NGS is reported to beat sampling (best-of-N), beam search, MCTS, and ACO on routing benchmarks, and to do so even when baselines are given a ten-fold larger sampling budget in several settings.
  • NGS maintains or improves performance under distribution shift, including TSPLib and CVRPLib instances that differ from the uniform random training distribution.
  • In molecular design, NGS serves as a learned alternative to hand-crafted GA crossover and mutation operators, improving average Top-10 scores over Graph GA, SMILES GA, STONED, and SynNet within the 10,000-evaluation budget.
  • Because NGS iteratively conditions on promising parents, it can replace conventional decoding in red-teaming and improves transfer to victim language models not seen during fine-tuning.

Reading between the lines

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

  • An implied next step is to use NGS as the loop inside training rather than only at test time, since the paper notes the policy can be fine-tuned while NGS runs; the same masking rule could bootstrap a weak pretrained policy into a stronger one.
  • A direct diagnostic is to measure how much probability mass the policy assigns to the parent vocabulary during generation; NGS should gain most when that mass is high and should approach plain sampling when it is near zero, which converts the paper's stated limitation into a testable prediction.
  • The same token-restriction mechanism could transfer to other sequential discrete-object domains the paper does not test, such as code repair or structured output generation, where parent-conditioned masking is equally cheap to implement.
  • A natural future comparison is NGS against self-improvement methods that interleave sampling with local search, since NGS's routing gains come without the handcrafted destruction rules those methods use.
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 / 5 minor

Summary. The paper proposes Neural Genetic Search (NGS), a test-time search method for sequential generative models. The core operator is a parent-conditioned crossover defined in Eq. (2), which restricts the next-token vocabulary to the union of tokens appearing in two selected parent sequences, with stochastic mutation (Eq. (3)) that occasionally removes this restriction. NGS maintains a population with rank-based selection and replacement (Eq. (5), Algorithm 1). The authors evaluate NGS on routing problems (TSP, CVRP, PCTSP, OP), red-teaming of language models, and de novo molecular design, reporting that NGS achieves smaller optimality gaps than sampling, beam search, MCTS, and ACO on routing, higher transfer toxicity than standard decoding schemes in red-teaming, and better average Top-10 scores than several molecular GAs. The paper claims NGS is a problem-agnostic, easy-to-implement decoding strategy that can be applied to any sequential generative model.

Significance. If the reported results hold, NGS is a genuinely simple and general idea: one token-restriction operator converts any sequential generative policy into an evolutionary search procedure, and the routing results in particular show large gains over strong baselines under the same pretrained heatmap, including at 10x lower budget. The paper is also commendable for controlling the policy across routing baselines, for providing code, and for explicitly acknowledging the dependence of the method on policy quality in the Limitations section. The empirical scope across three quite different domains supports the generality claim, though the red-teaming variant introduces task-specific modifications that qualify that claim.

major comments (4)
  1. [Section 3, Algorithm 1; Appendix E.3] The paper never isolates the effect of the token-restriction crossover, which is the stated core idea, from the effect of the population and rank-based elitist replacement. Algorithm 1 differs from ordinary sampling in three simultaneous ways: the parent-vocabulary mask (Eq. (2)), the population, and the rank-based selection/replacement (Eq. (5)). The sensitivity analysis in Appendix E.3 varies the mutation rate mu from 0.001 to 0.1 but never includes mu=1, which by Eq. (3) would reduce pNGS to p_theta* and disable the crossover mask while keeping the GA loop intact. Without this mutation-only control, the reported routing gains cannot be attributed to the crossover operator itself rather than to evolutionary population dynamics.
  2. [Section 5.2, Appendix A.2] The red-teaming version of NGS is not the generic Algorithm 1: Appendix A.2 adds two mechanisms, discarding used tokens from V_s1,s2 after each selection and replacing the reward rank with a novelty-weighted rank in Eq. (7). The paper states that the generic token restriction led to 'meaningless repetition' and that this behavior 'often hacks the reward function.' This directly qualifies the claim in Section 1 that NGS is problem-agnostic and applicable to any sequential generative model; the crossover as defined in Eq. (2) needed task-specific repair for this domain. The paper should either present these mechanisms as part of the proposed method (and test their necessity) or temper the universality claim.
  3. [Table 1, Section 5.1.2] The central routing results in Table 1 report no error bars, with the text stating that standard deviations are negligible. Since the main claim is that NGS 'achieves significantly smaller optimality gaps' than baselines, the main table should either include the standard deviations/confidence intervals or clearly refer to a table in the appendix that reports them; the reference to Figure 5 is not sufficient because Figure 5 is in Appendix E.1 and the main text does not report the actual values. Without variance information, the reader cannot assess whether the reported differences, some of which are small in absolute terms (e.g., CVRP N=500 NGS 1.840 vs ACO 1.975), are significant.
  4. [Table 4 and Appendix G] In the molecular design comparison, NGS uses 8K oracle calls to train a GFlowNet policy and 2K calls for the search, whereas the compared GAs (Graph GA, SMILES GA, STONED, SynNet) use the full 10K budget for search. This is within the PMO evaluation budget if training oracle calls are counted, but the comparison is no longer between test-time search procedures applied to a fixed pretrained policy; the neural policy has been reward-tuned on the same task distribution. The paper should state this asymmetry explicitly in the main text and should justify that the improvement over Graph GA is not primarily due to the trained policy rather than to the NGS search mechanism.
minor comments (5)
  1. [Appendix C.4] There is a typo: 'tshe reward' should read 'the reward.'
  2. [Section 5.2, Table 3] The Source column in Table 3 shows that NGS toxicity (0.71) is lower than several baselines (e.g., BS w=8 at 0.99, Temp tau=0.5 at 0.79); the text says NGS 'comparably balances' toxicity and diversity, but this should be stated more precisely, especially since the strongest NGS advantages appear only in the Transfer columns.
  3. [Appendix F, Table 7] In Table 7, where Llama-3.1-8B-Instruct is the source victim, NGS does not have the highest Source toxicity (0.65 vs Temp tau=0.5 at 0.73), and the paper's claim of significant outperformance should be tied to the transfer setting rather than stated as a universal property of the method.
  4. [Section 3.1, Eq. (2)] The notation V_s1,s2 is introduced as the union of tokens in the two parent sequences, but for routing problems the 'tokens' are edges and the union can contain infeasible edges; the paper explains this in the TSP example but the general notation in Eq. (2) would benefit from a remark that the mask is applied after any problem-specific feasibility masks, as is done in the concrete example.
  5. [Appendix B.1] The hyperparameter description says 'we use 100 for the number of ants in ACO and the number of offspring in NGS,' but it should clarify that the same mini-batch and iteration budget are used for both, which is stated only indirectly later in the same paragraph.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: NGS is a constructive decoding heuristic whose operator is defined by Eq. (2), and its performance claims are tested against external solvers with shared pretrained policies.

full rationale

The paper's derivation chain is constructive rather than circular. Eq. (1) defines the factored policy, Eq. (2) defines the token-restriction crossover as masked sampling from that policy, and Eq. (3) defines mutation as a mixture with the unmasked policy; Algorithm 1 wraps these definitions in a standard genetic loop. None of these equations is fitted to the reported outcomes, and none is derived from the claim that NGS outperforms other decoders. The routing and molecular policies come from the authors' prior work (Kim et al., 2025; Kim et al., 2024), but those citations supply a shared input policy and comparison baselines rather than the superiority result: all routing baselines use the same heatmap policy, and the reported gaps are measured against external references (Concorde, PyVRP, TSPLib, CVRPLib, and the PMO benchmark). The paper's own Limitations statement concedes that NGS depends on the quality of the underlying neural policy, which is an honest scope restriction, not a self-imported conclusion. The skeptic's concern that no μ=1 control isolates the mask's contribution is a legitimate empirical-attribution question, but it is not a circularity: the paper does not claim the mask's effect follows by construction from Eq. (2), and an omitted control does not make the operator's definition equivalent to its inputs. No step in the paper reduces a prediction to a fitted parameter or to an unverified self-citation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method has no mathematical derivation, so the ledger records hand-chosen algorithm hyperparameters, the trained policy as an input, the GFlowNet or supervised training recipe borrowed from prior work, and the red-teaming token-removal adjustment. The central claim is not built on fitted constants; the reported gains are measured against external solvers and prior GAs.

free parameters (5)
  • population size Npop = 100 (routing, molecules), 256 (red-teaming)
    Hand-chosen per task; sensitivity analysis in Figure 6 shows a modest effect on routing.
  • offspring size Noff = 100 (routing), 16 (red-teaming), 5 (molecules)
    Hand-chosen per task; Figure 6 shows larger Noff degrades routing performance.
  • stochastic mutation rate mu = 0.01 (routing, molecules), 0.05 (red-teaming)
    Hand-chosen; Figure 6 shows robustness across 0.001 to 0.1 in routing.
  • rank weight-shifting factor kappa = 0.001 (routing), 0.01 (red-teaming, molecules)
    Hand-chosen in Eq. (5); no sensitivity study is reported for kappa.
  • novelty rank weight omega = 0.1 (red-teaming only)
    Hand-chosen in Eq. (7) to balance toxicity and diversity; task-specific addition not present in the generic algorithm.
assumptions (4)
  • standard math The factorized decomposition p_theta(s) = prod p_theta(s_t | s_<t) (Eq. 1) holds and the policy is trained so that its conditional distributions are meaningful.
    The whole crossover is a reweighting of this decomposition; if the policy is untrained or mis-calibrated, Eq. (2) inherits the error.
  • domain assumption The pretrained policy assigns non-negligible probability to high-reward tokens within the parent vocabulary at most prefixes; otherwise crossover collapses to mutation and NGS reduces to sampling.
    Invoked in Eq. (2); the authors state in the Limitations that performance relies on the quality of the underlying neural policy.
  • domain assumption A reward function is available to score every generated candidate during search (full-information evaluation).
    Algorithm 1 evaluates r(s) for every offspring, which requires a cheap oracle; in molecular PMO this is the oracle, and in routing it is tour length.
  • ad hoc to paper For red-teaming, removing a token from V_s1,s2 after selection (A.2) and weighting rank by novelty (Eq. 7) prevents the reward hacking observed with the generic crossover.
    A.2 reports that the naive token restriction often hacks the reward function; the fix is task-specific and not part of the generic Algorithm 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Genetic Search in Discrete Spaces." pith.science (2026). https://pith.science/paper/ETCJDQD2

@misc{pith2026250210433,
  author       = {Pith},
  title        = {Pith review of: Neural Genetic Search in Discrete Spaces},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ETCJDQD2}},
  note         = {Machine review of arXiv:2502.10433}
}
read the original abstract

Effective search methods are crucial for improving the performance of deep generative models at test time. In this paper, we introduce a novel test-time search method, Neural Genetic Search (NGS), which incorporates the evolutionary mechanism of genetic algorithms into the generation procedure of deep models. The core idea behind NGS is its crossover, which is defined as parent-conditioned generation using trained generative models. This approach offers a versatile and easy-to-implement search algorithm for deep generative models. We demonstrate the effectiveness and flexibility of NGS through experiments across three distinct domains: routing problems, adversarial prompt generation for language models, and molecular design.

Figures

Figures reproduced from arXiv: 2502.10433 by the authors.

Figure 1
Figure 1. Illustrative examples of parents and offspring in various tasks. Offspring are generated through crossover and mutation. Crossover combines two parent chromosomes (Yellow and Blue) by restricting the vocabulary to tokens present in one of the parents. Mutation (Pink) occasionally removes this limitation, promoting the solution diversity. process to refine the outputs of generative models using existing problem-speci… view at source ↗
Figure 2
Figure 2. Overview of GA with NGS. (1) The pretrained generative policy sequentially constructs sequences, which correspond to chromosomes, to initialize the population. From this population, (2) parents are selected, and then (3) the policy reproduces offspring by sampling new sequences with a parent-conditioned mask. Finally, (4) the newly generated candidates replace members of the population, completing one evolutionary c… view at source ↗
Figure 3
Figure 3. An illustration of crossover and mutation in TSP (Section 3.1.1). At t = t ′ , the token-restriction crossover masks out edges not included in the parents, and then the pretrained policy gives the distribution over the remaining valid edges. At t = t ′ + 1, since all edges included in the parents are invalid ( ), the constraint-enforcing mutation activates, allowing selection of any unvisited node. 3.1.1. A CONCRETE… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Benchmark results on various routing problems. NGS outperformed sampling and ACO by a significant margin in all problems, showing its effectiveness as an inference-time search method. See Appendix E.1 for more comprehensive results. tional) probability distribution int…
Figure 5
Figure 5. Figure 5: shows extended results for the routing problems (Section 5.1), including the results for instances with 1,000 nodes. Overall, NGS substantially outperforms the baseline methods in all settings except for CVRP with 1,000 nodes. We suspect two reasons for the worse resul…
Figure 6
Figure 6. Figure 6: Sensitivity analysis on TSP and CVRP with 500 nodes. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

78 extracted references · 57 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    J., Javaheripi, M., Kauffmann, P., et al

    Abdin, M., Aneja, J., Behl, H., Bubeck, S., Eldan, R., Gunasekar, S., Harrison, M., Hewett, R. J., Javaheripi, M., Kauffmann, P., et al. Phi-4 technical report. arXiv preprint arXiv:2412.08905, 2024

  3. [3]

    Guiding deep molecular optimization with genetic exploration

    Ahn, S., Kim, J., Lee, H., and Shin, J. Guiding deep molecular optimization with genetic exploration. Advances in neural information processing systems (NeurIPS), 2020

  4. [4]

    Concorde TSP solver, 2006

    Applegate, D., Bixby, R., Chvatal, V., and Cook, W. Concorde TSP solver, 2006. URL https://www.math.uwaterloo.ca/tsp/concorde/

  5. [5]

    Constitutional AI : Harmlessness from AI feedback

    Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., Jones, A., Chen, A., Goldie, A., Mirhoseini, A., McKinnon, C., et al. Constitutional AI : Harmlessness from AI feedback. arXiv preprint arXiv:2212.08073, 2022

  6. [6]

    The prize collecting traveling salesman problem

    Balas, E. The prize collecting traveling salesman problem. Networks, 19 0 (6): 0 621--636, 1989

  7. [7]

    V., Norouzi, M., and Bengio, S

    Bello, I., Pham, H., Le, Q. V., Norouzi, M., and Bengio, S. Neural combinatorial optimization with reinforcement learning. arXiv preprint arXiv:1611.09940, 2016

  8. [8]

    Flow network based generative models for non-iterative diverse candidate generation

    Bengio, E., Jain, M., Korablyov, M., Precup, D., and Bengio, Y. Flow network based generative models for non-iterative diverse candidate generation. In Advances in Neural Information Processing Systems (NeurIPS), 2021

Show all 78 references
  1. [9]

    J., Tiwari, M., and Bengio, E

    Bengio, Y., Lahlou, S., Deleu, T., Hu, E. J., Tiwari, M., and Bengio, E. GFlowNet foundations. Journal of Machine Learning Research, 24 0 (210): 0 1--55, 2023

  2. [10]

    V., R \'e , C., and Mirhoseini, A

    Brown, B., Juravsky, J., Ehrlich, R., Clark, R., Le, Q. V., R \'e , C., and Mirhoseini, A. Large language monkeys: Scaling inference compute with repeated sampling. arXiv preprint arXiv:2407.21787, 2024

  3. [11]

    H., and Vaucher, A

    Brown, N., Fiscato, M., Segler, M. H., and Vaucher, A. C. GuacaMol : benchmarking models for de novo molecular design. Journal of Chemical Information and Modeling, 59 0 (3): 0 1096--1108, 2019

  4. [12]

    Evo P rompting: Language models for code-level neural architecture search

    Chen, A., Dohan, D., and So, D. Evo P rompting: Language models for code-level neural architecture search. Advances in neural information processing systems (NeurIPS), 2023 a

  5. [13]

    S., and Coates, M

    Chen, C., Zhang, Y., Fu, J., Liu, X. S., and Coates, M. Bidirectional learning for offline infinite-width model-based optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  6. [14]

    Bidirectional learning for offline model-based biological sequence design

    Chen, C., Zhang, Y., Liu, X., and Coates, M. Bidirectional learning for offline model-based biological sequence design. In International Conference on Machine Learning (ICML), 2023 b

  7. [15]

    Simulation-guided beam search for neural combinatorial optimization

    Choo, J., Kwon, Y.-D., Kim, J., Jae, J., Hottung, A., Tierney, K., and Gwon, Y. Simulation-guided beam search for neural combinatorial optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  8. [16]

    Dantzig, G. B. and Ramser, J. H. The truck dispatching problem. Management Science, 6 0 (1): 0 80--91, 1959

  9. [17]

    Genetic Algorithms and Robotics: A heuristic strategy for optimization, volume 1

    Davidor, Y. Genetic Algorithms and Robotics: A heuristic strategy for optimization, volume 1. World Scientific Publishing Company, 1991

  10. [18]

    A survey on LLM inference-time self-improvement

    Dong, X., Teleki, M., and Caverlee, J. A survey on LLM inference-time self-improvement. arXiv preprint arXiv:2412.14352, 2024

  11. [19]

    Reinforced genetic algorithm for structure-based drug design

    Fu, T., Gao, W., Coley, C., and Sun, J. Reinforced genetic algorithm for structure-based drug design. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  12. [20]

    Sample efficiency matters: a benchmark for practical molecular optimization

    Gao, W., Fu, T., Sun, J., and Coley, C. Sample efficiency matters: a benchmark for practical molecular optimization. Advances in Neural Information Processing Systems (NeurIPS), 2022 a

  13. [21]

    Gao, W., Mercado, R., and Coley, C. W. Amortized tree generation for bottom-up synthesis planning and synthesizable molecular design. In International Conference on Learning Representations (ICLR), 2022 b

  14. [22]

    Gao, W., Luo, S., and Coley, C. W. Generative artificial intelligence for navigating synthesizable chemical space. arXiv preprint arXiv:2410.03494, 2024

  15. [23]

    L., Levy, L., and Vohra, R

    Golden, B. L., Levy, L., and Vohra, R. The orienteering problem. Naval Research Logistics (NRL), 34 0 (3): 0 307--318, 1987

  16. [24]

    F., de Magalhães Mendes , J

    Gonçalves, J. F., de Magalhães Mendes , J. J., and Resende, M. G. A hybrid genetic algorithm for the job shop scheduling problem. European Journal of Operational Research, 167 0 (1): 0 77--95, 2005. ISSN 0377-2217. doi:https://doi.org/10.1016/j.ejor.2004.03.012. URL https://ww...

  17. [25]

    Connecting large language models with evolutionary algorithms yields powerful prompt optimizers

    Guo, Q., Wang, R., Guo, J., Li, B., Song, K., Tan, X., Liu, G., Bian, J., and Yang, Y. Connecting large language models with evolutionary algorithms yields powerful prompt optimizers. In International Conference on Learning Representations (ICLR), 2024

  18. [26]

    An extension of the Lin-Kernighan-Helsgaun TSP solver for constrained traveling salesman and vehicle routing problems

    Helsgaun, K. An extension of the Lin-Kernighan-Helsgaun TSP solver for constrained traveling salesman and vehicle routing problems. Roskilde: Roskilde University, pp.\ 966--980, 12 2017. doi:10.13140/RG.2.2.25569.40807

  19. [27]

    Holland, J. H. Adaptation in natural and artificial systems: an introductory analysis with applications to biology, control, and artificial intelligence. MIT press, 1992

  20. [28]

    The curious case of neural text degeneration

    Holtzman, A., Buys, J., Du, L., Forbes, M., and Choi, Y. The curious case of neural text degeneration. In International Conference on Learning Representations (ICLR), 2020

  21. [29]

    Efficient active search for combinatorial optimization problems

    Hottung, A., Kwon, Y.-D., and Tierney, K. Efficient active search for combinatorial optimization problems. In International Conference on Learning Representations (ICLR), 2022

  22. [30]

    A mobile robot path planning using genetic algorithm in static environment

    Ismail, A., Sheta, A., and Al-Weshah, M. A mobile robot path planning using genetic algorithm in static environment. Journal of Computer Science, 4 0 (4): 0 341--344, 2008

  23. [31]

    Jensen, J. H. A graph-based genetic algorithm and generative model/ Monte Carlo tree search for the exploration of chemical space. Chemical Science, 10 0 (12): 0 3567--3572, 2019

  24. [32]

    K., Cappart, Q., Rousseau, L.-M., and Laurent, T

    Joshi, C. K., Cappart, Q., Rousseau, L.-M., and Laurent, T. Learning TSP requires rethinking generalization. In International Conference on Principles and Practice of Constraint Programming (CP), 2021

  25. [33]

    and De Winter, H

    Kerstjens, A. and De Winter, H. LEADD : Lamarckian evolutionary algorithm for de novo drug design. Journal of Cheminformatics, 14 0 (1): 0 3, 2022

  26. [34]

    Genetic-guided GFlowNets for sample efficient molecular optimization

    Kim, H., Kim, M., Choi, S., and Park, J. Genetic-guided GFlowNets for sample efficient molecular optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2024

  27. [35]

    Bootstrapped training of score-conditioned generator for offline design of biological sequences

    Kim, M., Berto, F., Ahn, S., and Park, J. Bootstrapped training of score-conditioned generator for offline design of biological sequences. In Advances in Neural Information Processing Systems (NeurIPS), 2023

  28. [36]

    Ant colony sampling with GFlowNets for combinatorial optimization

    Kim, M., Choi, S., Son, J., Kim, H., Park, J., and Bengio, Y. Ant colony sampling with GFlowNets for combinatorial optimization. International Conference on Artificial Intelligence and Statistics (AISTATS), 2025

  29. [37]

    Kobeaga, G., Merino, M., and Lozano, J. A. An efficient evolutionary algorithm for the orienteering problem. Computers & Operations Research, 90: 0 42--59, 2018. ISSN 0305-0548. doi:https://doi.org/10.1016/j.cor.2017.09.003. URL https://www.sciencedirect.com/science/article/pi...

  30. [38]

    Attention, learn to solve routing problems! In International Conference on Learning Representations (ICLR), 2019

    Kool, W., van Hoof, H., and Welling, M. Attention, learn to solve routing problems! In International Conference on Learning Representations (ICLR), 2019

  31. [39]

    Deep policy dynamic programming for vehicle routing problems

    Kool, W., van Hoof, H., Gromicho, J., and Welling, M. Deep policy dynamic programming for vehicle routing problems. In International Conference on Integration of Constraint Programming, Artificial Intelligence, and Operations Research (CPAIOR), 2022

  32. [40]

    Self-referencing embedded strings ( SELFIES ): A 100\ Machine Learning: Science and Technology, 1 0 (4): 0 045024, 2020

    Krenn, M., H \"a se, F., Nigam, A., Friederich, P., and Aspuru-Guzik, A. Self-referencing embedded strings ( SELFIES ): A 100\ Machine Learning: Science and Technology, 1 0 (4): 0 045024, 2020

  33. [41]

    Genetic algorithm based approach for autonomous mobile robot path planning

    Lamini, C., Benhlima, S., and Elbekri, A. Genetic algorithm based approach for autonomous mobile robot path planning. Procedia Computer Science, 127: 0 180--189, 2018

  34. [42]

    J., Kawaguchi, K., Gidel, G., Bengio, Y., Malkin, N., et al

    Lee, S., Kim, M., Cherif, L., Dobre, D., Lee, J., Hwang, S. J., Kawaguchi, K., Gidel, G., Bengio, Y., Malkin, N., et al. Learning diverse attacks on large language models for robust red-teaming and safety tuning. arXiv preprint arXiv:2405.18540, 2024 a

  35. [43]

    P., Liu, M., Reidenbach, D., Paliwal, S

    Lee, S., Kreis, K., Veccham, S. P., Liu, M., Reidenbach, D., Paliwal, S. G., Vahdat, A., and Nie, W. Molecule generation with fragment retrieval augmentation. In Advances in Neural Information Processing Systems (NeurIPS), 2024 b

  36. [44]

    Lehman, J., Gordon, J., Jain, S., Ndousse, K., Yeh, C., and Stanley, K. O. Evolution through large models. In Handbook of Evolutionary Machine Learning, pp.\ 331--366. Springer, 2023

  37. [45]

    Auto DAN : Generating stealthy jailbreak prompts on aligned large language models

    Liu, X., Xu, N., Chen, M., and Xiao, C. Auto DAN : Generating stealthy jailbreak prompts on aligned large language models. In International Conference on Learning Representations (ICLR), 2024

  38. [46]

    Llama Team, A. . M. The llama 3 herd of models, 2024. URL https://arxiv.org/abs/2407.21783

  39. [47]

    Neural combinatorial optimization with heavy decoder: Toward large scale generalization

    Luo, F., Lin, X., Liu, F., Zhang, Q., and Wang, Z. Neural combinatorial optimization with heavy decoder: Toward large scale generalization. In Advances in Neural Information Processing Systems (NeurIPS), 2023

  40. [48]

    P., Hermann, K., Welleck, S., Yazdanbakhsh, A., and Clark, P

    Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y., Gupta, S., Majumder, B. P., Hermann, K., Welleck, S., Yazdanbakhsh, A., and Clark, P. Self-refine: Iterative refinement with self-feedback. In Advances in Ne...

  41. [49]

    and Kwon, C

    Mahmoudinazlou, S. and Kwon, C. A hybrid genetic algorithm for the min--max multiple traveling salesman problem. Computers & Operations Research, 162: 0 106455, 2024

  42. [50]

    GFlowNets and variational inference

    Malkin, N., Lahlou, S., Deleu, T., Ji, X., Hu, E., Everett, K., Zhang, D., and Bengio, Y. GFlowNets and variational inference. In International Conference on Learning Representations (ICLR), 2023

  43. [51]

    J., Bradley, H., Gaier, A., Moradi, A., Hoover, A

    Meyerson, E., Nelson, M. J., Bradley, H., Gaier, A., Moradi, A., Hoover, A. K., and Lehman, J. Language model crossover: Variation through few-shot prompting. ACM Transactions on Evolutionary Learning, 4 0 (4): 0 1--40, 2024

  44. [52]

    M., Goodsell, D

    Morris, G. M., Goodsell, D. S., Halliday, R. S., Huey, R., Hart, W. E., Belew, R. K., and Olson, A. J. Automated docking using a lamarckian genetic algorithm and an empirical binding free energy function. Journal of computational chemistry, 19 0 (14): 0 1639--1662, 1998

  45. [53]

    Genetic algorithms for flowshop scheduling problems

    Murata, T., Ishibuchi, H., and Tanaka, H. Genetic algorithms for flowshop scheduling problems. Computers & Industrial Engineering, 30 0 (4): 0 1061--1071, 1996. ISSN 0360-8352. doi:https://doi.org/10.1016/0360-8352(96)00053-8. URL https://www.sciencedirect.com/science/article/...

  46. [54]

    and Kobayashi, S

    Nagata, Y. and Kobayashi, S. A powerful genetic algorithm using edge assembly crossover for the traveling salesman problem. INFORMS Journal on Computing, 25 0 (2): 0 346--363, 2013

  47. [55]

    Beyond generative models: superfast traversal, optimization, novelty, exploration and discovery (STONED) algorithm for molecules using SELFIES

    Nigam, A., Pollice, R., Krenn, M., dos Passos Gomes, G., and Aspuru-Guzik, A. Beyond generative models: superfast traversal, optimization, novelty, exploration and discovery (STONED) algorithm for molecules using SELFIES . Chemical Science, 12 0 (20): 0 7079--7090, 2021

  48. [56]

    Molecular de-novo design through deep reinforcement learning

    Olivecrona, M., Blaschke, T., Engkvist, O., and Chen, H. Molecular de-novo design through deep reinforcement learning. Journal of Cheminformatics, 9 0 (1): 0 1--14, 2017

  49. [57]

    Omara, F. A. and Arafa, M. M. Genetic algorithms for task scheduling problem. Journal of Parallel and Distributed Computing, 70 0 (1): 0 13--22, 2010. ISSN 0743-7315. doi:https://doi.org/10.1016/j.jpdc.2009.09.009. URL https://www.sciencedirect.com/science/article/pii/S0743731...

  50. [58]

    Red teaming language models with language models

    Perez, E., Huang, S., Song, F., Cai, T., Ring, R., Aslanides, J., Glaese, A., McAleese, N., and Irving, G. Red teaming language models with language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2022

  51. [59]

    DIMES : A differentiable meta solver for combinatorial optimization problems

    Qiu, R., Sun, Z., and Yang, Y. DIMES : A differentiable meta solver for combinatorial optimization problems. In Advances in Neural Information Processing Systems (NeurIPS), 2022

  52. [60]

    Language models are unsupervised multitask learners

    Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., Sutskever, I., et al. Language models are unsupervised multitask learners. OpenAI blog, 1 0 (8): 0 9, 2019

  53. [61]

    J., Geng, X., Herschl, M

    Reddy, A. J., Geng, X., Herschl, M. H., Kolli, S., Kumar, A., Hsu, P. D., Levine, S., and Ioannidis, N. M. Designing cell-type-specific promoter sequences using conservative model-based optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2024

  54. [62]

    TSPLIB—A traveling salesman problem library

    Reinelt, G. TSPLIB—A traveling salesman problem library. ORSA journal on computing, 3 0 (4): 0 376--384, 1991

  55. [63]

    Scaling llm test-time compute optimally can be more effective than scaling model parameters

    Snell, C., Lee, J., Xu, K., and Kumar, A. Scaling llm test-time compute optimally can be more effective than scaling model parameters. arXiv preprint arXiv:2408.03314, 2024

  56. [64]

    Meta-SAGE : scale meta-learning scheduled adaptation with guided exploration for mitigating scale shift on combinatorial optimization

    Son, J., Kim, M., Kim, H., and Park, J. Meta-SAGE : scale meta-learning scheduled adaptation with guided exploration for mitigating scale shift on combinatorial optimization. In International Conference on Machine Learning (ICML), 2023

  57. [65]

    G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al

    Team, G., Riviere, M., Pathak, S., Sessa, P. G., Hardin, C., Bhupatiraju, S., Hussenot, L., Mesnard, T., Shahriari, B., Ram \'e , A., et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024

  58. [66]

    Conservative objective models for effective offline model-based optimization

    Trabucco, B., Kumar, A., Geng, X., and Levine, S. Conservative objective models for effective offline model-based optimization. In International Conference on Machine Learning (ICML), 2021

  59. [67]

    Tripp, A., Daxberger, E., and Hern \'a ndez-Lobato, J. M. Sample-efficient optimization in the latent space of deep generative models via weighted retraining. Advances in Neural Information Processing Systems (NeurIPS), 2020

  60. [68]

    New benchmark instances for the capacitated vehicle routing problem

    Uchoa, E., Pecin, D., Pessoa, A., Poggi, M., Vidal, T., and Subramanian, A. New benchmark instances for the capacitated vehicle routing problem. European Journal of Operational Research, 257 0 (3): 0 845--858, 2017

  61. [69]

    Hybrid genetic search for the cvrp: Open-source implementation and swap* neighborhood

    Vidal, T. Hybrid genetic search for the cvrp: Open-source implementation and swap* neighborhood. Computers & Operations Research, 140: 0 105643, 2022

  62. [70]

    G., Gendreau, M., Lahrichi, N., and Rei, W

    Vidal, T., Crainic, T. G., Gendreau, M., Lahrichi, N., and Rei, W. A hybrid genetic algorithm for multidepot and periodic vehicle routing problems. Operations Research, 60 0 (3): 0 611--624, 2012

  63. [71]

    Efficient evolutionary search over chemical space with large language models

    Wang, H., Skreta, M., Ser, C.-T., Gao, W., Kong, L., Strieth-Kalthoff, F., Duan, C., Zhuang, Y., Yu, Y., Zhu, Y., et al. Efficient evolutionary search over chemical space with large language models. arXiv preprint arXiv:2406.16976, 2024

  64. [72]

    Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers

    Wang, W., Bao, H., Huang, S., Dong, L., and Wei, F. Minilmv2: Multi-head self-attention relation distillation for compressing pretrained transformers. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pp.\ 2140--2151, 2021

  65. [73]

    SMILES , a chemical language and information system

    Weininger, D. SMILES , a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of chemical information and computer sciences, 28 0 (1): 0 31--36, 1988

  66. [74]

    A., Lan, L., and Kool, W

    Wouda, N. A., Lan, L., and Kool, W. PyVRP : a high-performance VRP solver package. INFORMS Journal on Computing, 2024. doi:10.1287/ijoc.2023.0055. URL https://doi.org/10.1287/ijoc.2023.0055

  67. [75]

    Qwen2 technical report

    Yang, A., Yang, B., Hui, B., Zheng, B., Yu, B., Zhou, C., Li, C., Li, C., Liu, D., Huang, F., Dong, G., Wei, H., Lin, H., Tang, J., Wang, J., Yang, J., Tu, J., Zhang, J., Ma, J., Xu, J., Zhou, J., Bai, J., He, J., Lin, J., Dang, K., Lu, K., Chen, K., Yang, K., Li, M., Xue, M.,...

  68. [76]

    DeepACO : Neural-enhanced ant systems for combinatorial optimization

    Ye, H., Wang, J., Cao, Z., Liang, H., and Li, Y. DeepACO : Neural-enhanced ant systems for combinatorial optimization. In Advances in Neural Information Processing Systems (NeurIPS), 2023

  69. [77]

    Re E vo: Large language models as hyper-heuristics with reflective evolution

    Ye, H., Wang, J., Cao, Z., Berto, F., Hua, C., Kim, H., Park, J., and Song, G. Re E vo: Large language models as hyper-heuristics with reflective evolution. In Advances in Neural Information Processing Systems (NeurIPS), 2024

  70. [78]

    RoMA : Robust model adaptation for offline model-based optimization

    Yu, S., Ahn, S., Song, L., and Shin, J. RoMA : Robust model adaptation for offline model-based optimization. Advances in Neural Information Processing Systems (NeurIPS), 2021

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.