Pith. sign in

REVIEW 4 major objections 4 minor 54 references

An LLM-Empowered Adaptive Evolutionary Algorithm For Multi-Component Deep Learning Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read The paper claims that using a large language model to seed and re-seed an evolutionary search makes safety-violation discovery in multi-component deep learning systems more efficient and more diverse.

desk verdict LLM-seeded evolutionary search for AV safety violations is a plausible idea, but the evaluation confound between prompt-injected domain knowledge and the adaptive algorithm leaves the central claim unproven. read the letter →

arxiv 2501.00829 v2 pith:TOTKO4QL submitted 2025-01-01 cs.NE cs.AI

Haoxiang Tian , Xingshuo Han , Guoquan Wu , An Guo , Yuan Zhou. Jie Zhang , Shuo Li , Jun Wei , Tianwei Zhang This is my paper · ORCID
classification cs.NEcs.AI
keywords largelanguagemodelsmulti-objectiveevolutionaryalgorithmsafetyviolationdetectionautonomousdrivingtestscenariogenerationdifferentialseedsadaptivevariation
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 an evolutionary search for safety violations in multi-component deep learning systems becomes both more efficient and more diverse when a large language model (LLM) is used to seed the initial population and to generate 'differential seeds' when the search stalls. Tested on the Apollo autonomous driving stack in the SORA-SVL simulator, the proposed µMOEA finds ten types of safety violations in 24 hours, whereas the state-of-the-art MOEA baseline MOSAT finds six. It also needs on average 12 generated solutions per detected violation versus 62.1 for MOSAT, and finds its first violation in about 7 minutes versus 16. The point of the claim is that LLM-driven guidance can replace blind random initialization and help escape local optima, making failure discovery for complex systems substantially cheaper.

What carries the argument

The central object is µMOEA, a three-part evolutionary loop: an LLM-generated initial population, adaptive selection and variation with dynamically computed crossover and mutation probabilities and per-chromosome variation types, and LLM-generated differential seeds produced from evolutionary feedback. Differential seeds are new solutions the LLM creates after the search stagnates, constrained to differ from previously seen solutions by a computed Euclidean distance threshold. The machinery also includes the two fitness objectives: criticality, the minimal distance of a generated scenario to a safety violation, and diversity, the mean Euclidean distance from a scenario to all others in the population. This combination is what the paper argues carries the efficiency and diversity gains.

What would settle it

Run the baseline MOSAT with the same scenario-structure hints that the µMOEA prompts contain—e.g., multi-lane trajectories that disturb the ego vehicle—encoded as constraints on the random initialization, and compare results over the same 24-hour budget. If the gap in violation types and scenario count closes, the LLM's injected domain knowledge, not the adaptive operators, is doing the work.

Watch

Extended reading notes

Core claim

In the paper's own framing, the central discovery is that the two bottlenecks of MOEA-based safety testing—random initialization and premature convergence to similar solutions—can both be addressed by prompting an LLM. For initialization, µMOEA gives the LLM a description of the solution structure, parameter ranges, and examples of desirable test cases, so the first generation already targets the objectives. For convergence, when the elite solutions have not changed for a fixed number of generations, µMOEA feeds the best and worst chromosomes back into the LLM, which then generates differential seeds that are intentionally separated from the current optimum by a minimum Euclidean distance, with a re-prompt if the seeds are not different enough. The adaptive crossover and mutation operators then vary the mixture of exploration and exploitation generation by generation. The paper reports that this combined mechanism discovers ten distinct types of safety violations in Apollo, all within the first 14 hours, and that the six types found by the baseline are a strict subset.

Load-bearing premise

The improvement is attributed to µMOEA's algorithm, but its prompts inject scenario-structure knowledge that the baseline lacks, and the LLM may already know the test environment from its training data, so the measured gains could come from hidden domain knowledge rather than from the adaptive search itself.

Editorial extensions

If this is right

  • A 24-hour safety run over Apollo can uncover ten violation types instead of six, so the same compute budget buys roughly 67% more failure classes.
  • The first safety violation appears in roughly 7 minutes on average, which makes LLM-assisted search practical for continuous integration loops where early feedback matters.
  • The strategy of LLM seeding and re-seeding is not tied to a specific MOEA backbone; it is described as a generic wrapper around adaptive selection and variation, so it could be ported to other Pareto-based algorithms.
  • The method's performance depends on prompt quality and the LLM's quantitative reasoning, and the authors note that remote API latency and token limits are current constraints.

Reading between the lines

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

  • The prompt template shown in Table 1 encodes scenario-structure knowledge—for example, that trajectories should involve different lanes and should disturb the ego vehicle—so part of the gain over MOSAT may come from this injected expertise rather than from the adaptive operators; a fairer test would hand the same hints to the baseline as constraints.
  • GPT-4 has likely been trained on public descriptions of Apollo and San Francisco road networks, so the LLM's 'initial population' and 'differential seeds' may include memorized driving scenarios; using a local model without that exposure would test whether the gains are general.
  • The token-limit workaround of incrementally updating feedback prompts could be improved by actively selecting which historical solutions to show the LLM, which the authors list as future work and might further raise diversity.
  • The same seeding and re-seeding pattern could transfer to other test-generation settings where a language model can encode domain constraints, e.g., fuzzing for network protocol implementations or scenario search in other robotics simulators.
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 / 4 minor

Summary. The paper proposes µMOEA, an LLM-empowered adaptive evolutionary algorithm for detecting safety violations in multi-component deep learning systems. Instead of random initialization, µMOEA uses GPT-4 to generate an initial population from a domain-structured prompt, then applies adaptive crossover/mutation probabilities that depend on population fitness statistics, and periodically feeds the evolutionary history back to the LLM to generate 'differential seeds' that escape local optima. The method is evaluated on the Apollo autonomous driving system with a San Francisco map over 24-hour runs, using MOSAT (an NSGA-II-based method) as the baseline. The reported results claim that µMOEA finds more types of safety violations (10 vs. 6), needs fewer generated solutions per violation (12 vs. 62.1), and detects the first violation faster (7 vs. 16 minutes on average).

Significance. If the central claim is established, the paper would demonstrate a practically valuable way to combine LLM priors with evolutionary search: cheaper and more diverse discovery of safety violations in a realistic industrial ADS. The architecture is clear and the choice of an industrial full-stack ADS with 24-hour runs is a useful stress test. However, the current evaluation does not yet establish the claim because the comparison is confounded by domain knowledge injected through the LLM prompts, the statistical basis is thin (five runs, no variance or significance tests, one map, one system), and the safety-violation-type taxonomy is produced by the authors' own manual root-cause analysis without a documented rubric. The paper also does not release code or data, limiting reproducibility. The algorithmic contribution is interesting and plausible, but the load-bearing evidence needs strengthening before the efficiency and diversity claims can be accepted.

major comments (4)
  1. [Methodology, 'Instructing LLM to Create Initial Population'; Evaluation, Table 4] The comparison with MOSAT is confounded by scenario-structure knowledge embedded in the LLM prompts. The prompt template shown in Table 1 requires NPC trajectories to 'disturb the ego vehicle's driving path' and to 'involve different lanes', while the evaluation uses Apollo on a San Francisco map; MOSAT's random initialization receives no such hints, and GPT-4 may have memorized Apollo or San Francisco driving scenarios during pretraining. The ablations in Table 3 do not remove this confound: µM OEAr keeps the LLM-based differential seeds and µM OEAn keeps the LLM-generated initial population, so neither variant is a non-LLM version with the same adaptive operators and no domain priors. Consequently, the gains in Table 4 (10 vs. 6 violation types; 12 vs. 62.1 solutions per violation) could be due to injected or memorized domain knowledge rather than to the adaptive selection, adaptive variation, or differential-seed mechanism. I request a controlled ablation with random initialization plus the same adaptive operators and no LLM knowledge, and ideally a variant whose LLM prompt only encodes neutral, domain-agnostic structure.
  2. [Evaluation, Tables 3 and 4] The central efficiency and diversity claims rest on averages over five runs, reported without standard deviations or significance tests. Table 4 reports only min/max and average values, and Table 3 gives no variance information at all; with one ADS, one map, and one baseline, the differences (e.g., 10 vs. 6 types, 7 vs. 16 minutes) could be within run-to-run variability. Please report per-run results, effect sizes, and appropriate significance tests, or explicitly temper the 'significantly improve' claim in the abstract and conclusion.
  3. [Evaluation, 'Effectiveness of µMOEA'] The diversity metric that drives the main claim is the number of safety-violation types, but these types come from the authors' manual root-cause analysis and no classification rubric, inter-rater agreement, or independent validation is provided. Because the 10-vs-6 comparison is the headline result, the taxonomy must be reproducible; otherwise the diversity advantage cannot be independently verified.
  4. [Methodology, 'Adaptive Selection'] The equations for the adaptive crossover and mutation probabilities are not self-consistent. The text defines P_Mi as the mutation probability of s_i, yet the formula in the crossover paragraph is labeled P_Mc and contains k2 and k4; the constant k1 and k3 are later assigned values but do not appear in either formula. In addition, the denominators (fmax - fbar) and (fbar - fmin) are zero when the population's fitness values coincide, and the degenerate case is not handled. These issues make the core adaptive mechanism difficult to reproduce exactly, so the presentation needs to be corrected.
minor comments (4)
  1. [Methodology, Algorithm 1] The pseudocode uses nonstandard or corrupted notation such as 'P ← P S LLM generate', 'SCR ← SCR S pi', 'P N ← P N S {p′}', and 'TS, SC, MS ← ∅'; these should be written with explicit set-union symbols and clear line references, since the text refers to lines 2 and 13-14 that do not match the listed algorithm lines.
  2. [Evaluation, 'Effectiveness of µMOEA' and 'Advancement of µMOEA'] The reported average Euclidean distance for µM OEAr is 69.72 meters in the effectiveness subsection but 64.70 meters in the advancement subsection; please reconcile this inconsistency and state which value corresponds to which set of runs.
  3. [Evaluation, Tables 3 and 4] The time to detect all safety-violation types for µMOEA is reported as 14 hours in Table 3 and as an average of 12.9 hours in Table 4, while the text says all types are revealed in the first 14 hours; these numbers need to be reconciled.
  4. [Throughout] There are several typos and formatting issues: Table 4's header reads 'MOSA T' instead of 'MOSAT', Figure 2's caption reads 'safty violations', and the MOSAT description in the advancement subsection contains 'csearch' instead of 'to search'. Additionally, no code or data are released, which makes the numerical results impossible to reproduce; consider providing an artifact or a detailed replication package.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's contribution is an empirically evaluated algorithm, and its reported gains are measured outcomes rather than restatements of its inputs.

full rationale

I examined the claimed derivation chain. The paper does not derive a prediction from fitted parameters; it proposes an algorithm and measures its outputs on an external industrial system (Baidu Apollo via SORA-SVL). The adaptive selection and variation equations depend only on current-population fitness values and user-set constants; no quantity that is reported as a result (types of safety violations, time-to-first-violation, solutions per violation) appears as an input to the fitness function or to the LLM prompts. The differential-seed mechanism is a genuine additional loop: the LLM is asked to produce seeds different from previous solutions, and the difference check filters those seeds, but the paper does not claim this check itself predicts the measured diversity; the measured diversity is computed after execution in the simulator. The comparison baseline MOSAT is the authors' prior work, but it is used as an external empirical baseline, not as a justification that entails the conclusion; the comparison is therefore a self-citation in the ordinary benchmarking sense, not a load-bearing circular argument. The strongest legitimate concern is evaluation fairness: the LLM prompts encode domain knowledge (for example, requiring NPC trajectories to disturb the ego vehicle and to involve different lanes), so part of the measured advantage over random initialization could come from injected priors rather than from the adaptive evolutionary operators. That is a validity risk, not a definitional circularity, because the reported values are not constructed to equal the prompts or the fitness values. Under the hard rule requiring an explicit reduction of an output to an input, no circular step is exhibited.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim is an empirical efficiency claim, so it rests on the experimental setup: one simulator, one map, one LLM API, one baseline, and hand-chosen search constants. These assumptions and parameters, rather than a formal derivation, are what support the reported 10-versus-6 and 12-versus-62 results.

free parameters (6)
  • k1, k2 = 0.6
    Hand-chosen constants in the adaptive selection probability formulas; no sensitivity analysis is provided.
  • k3, k4 = 1.0
    Hand-chosen caps on crossover and mutation probabilities; no sensitivity analysis is provided.
  • c1, m1 (threshold coefficients) = not reported
    Coefficients in the threshold^n formula that control when mutation and crossover are applied; values are not given and appear to be chosen by hand.
  • stagnation window t = not reported
    Number of consecutive generations with unchanged high-fitness solutions that triggers LLM differential seed generation; this is central to the C2 mechanism but is unspecified.
  • population size N = not reported
    Number of individuals per generation; the paper only says it matches MOSAT's population size.
  • LLM prompt templates and feedback rules = hand-crafted
    The exact prompts in Tables 1 and 2 are illustrative patterns, not the literal strings used; the content encodes domain knowledge and is central to the method.
assumptions (5)
  • domain assumption SORA-SVL execution is a valid proxy for Apollo safety behavior.
    The paper runs generated scenarios in SORA-SVL with a San Francisco map and treats simulator outcomes as safety violations; no validation against real-world Apollo behavior is provided.
  • domain assumption The criticality and diversity metrics f_c and f_d capture the objectives of interest.
    The safety-violation distance and trajectory Euclidean distance are used as fitness, but no evidence shows they correlate with actual dangerous outcomes.
  • domain assumption LLM outputs can be parsed into executable scenarios that respect value ranges.
    The method relies on GPT-4 producing well-formed solutions from prompts without a validation or repair step beyond difference checking.
  • domain assumption MOSAT is a fair state-of-the-art baseline.
    MOSAT is the authors' own prior method and is the only comparison; no external re-implementation or additional MOEA baselines are used.
  • standard math NSGA-II's non-dominated sorting and crowding distance are correct and appropriate.
    The adaptive operators build on the NSGA-II framework cited from Deb et al. 2002.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An LLM-Empowered Adaptive Evolutionary Algorithm For Multi-Component Deep Learning Systems." pith.science (2026). https://pith.science/paper/TOTKO4QL

@misc{pith2026250100829,
  author       = {Pith},
  title        = {Pith review of: An LLM-Empowered Adaptive Evolutionary Algorithm For Multi-Component Deep Learning Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TOTKO4QL}},
  note         = {Machine review of arXiv:2501.00829}
}
abstract

Multi-objective evolutionary algorithms (MOEAs) are widely used for searching optimal solutions in complex multi-component applications. Traditional MOEAs for multi-component deep learning (MCDL) systems face challenges in enhancing the search efficiency while maintaining the diversity. To combat these, this paper proposes $\mu$MOEA, the first LLM-empowered adaptive evolutionary search algorithm to detect safety violations in MCDL systems. Inspired by the context-understanding ability of Large Language Models (LLMs), $\mu$MOEA promotes the LLM to comprehend the optimization problem and generate an initial population tailed to evolutionary objectives. Subsequently, it employs adaptive selection and variation to iteratively produce offspring, balancing the evolutionary efficiency and diversity. During the evolutionary process, to navigate away from the local optima, $\mu$MOEA integrates the evolutionary experience back into the LLM. This utilization harnesses the LLM's quantitative reasoning prowess to generate differential seeds, breaking away from current optimal solutions. We evaluate $\mu$MOEA in finding safety violations of MCDL systems, and compare its performance with state-of-the-art MOEA methods. Experimental results show that $\mu$MOEA can significantly improve the efficiency and diversity of the evolutionary search.

Figures

Figures reproduced from arXiv: 2501.00829 by the authors.

Figure 1
Figure 1. It consists of three steps: instructing the LLM to create the initial population (line 2), evolving the popula￾tion adaptively to search for optimal solutions (line 3-12), guiding the LLM to generate differential seeds based on the feedback of the evolutionary process (line 13-14). Below we give detailed explanation of each step. Instructing LLM to Create Initial Population When starting an evolutionary search, inst… view at source ↗
Figure 2
Figure 2. The number of found safety violations over time [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

54 extracted references · 32 canonical work pages

  1. [1]

    Microbial Genetic Algorithm-based Black-box Attack against Interpretable Deep Learning Systems

    Abdukhamidov, E.; Abuhamad, M.; Woo, S. S.; Chan-Tin, E.; and Abuhmed, T. 2023 a . Microbial Genetic Algorithm-based Black-box Attack against Interpretable Deep Learning Systems. arXiv preprint arXiv:2307.06496

    work page Pith review arXiv 2023

  2. [2]

    Unveiling Vulnerabilities in Interpretable Deep Learning Systems with Query-Efficient Black-box Attacks

    Abdukhamidov, E.; Abuhamad, M.; Woo, S. S.; Chan-Tin, E.; and Abuhmed, T. 2023 b . Unveiling Vulnerabilities in Interpretable Deep Learning Systems with Query-Efficient Black-box Attacks. arXiv preprint arXiv:2307.11906

    work page Pith review arXiv 2023

  3. [3]

    Amodei, D.; Olah, C.; Steinhardt, J.; Christiano, P.; Schulman, J.; and Man \'e , D. 2016. Concrete problems in AI safety. arXiv preprint arXiv:1606.06565

    arXiv 2016

  4. [4]

    Asharf, J.; Moustafa, N.; Khurshid, H.; Debie, E.; Haider, W.; and Wahab, A. 2020. A review of intrusion detection systems using machine and deep learning in internet of things: Challenges, solutions and future directions. Electronics, 9(7): 1177

    work page 2020

  5. [5]

    Autoware Foundation . 2015. Autoware Overview. https://autoware.org/autoware-overview/. Accessed: 2023-09-01

    work page 2015

  6. [6]

    April 1, 2024

    Baidu, I. April 1, 2024. Baidu Launches Public Robotaxi Trial Operation

    work page 2024

  7. [7]

    Baidu Apollo . 2013. An open autonomous driving platform. https://github.com/ApolloAuto/apollo. Accessed: 2022-03-16

    work page 2013

  8. [8]

    C.; and Stifter, T

    Ben Abdessalem, R.; Nejati, S.; Briand, L. C.; and Stifter, T. 2016. Testing advanced driver assistance systems using multi-objective search and neural networks. In Proceedings of the 31st IEEE/ACM International Conference on Automated Software Engineering, 63--74. Doi: 10.1145/2970276.2970311

    arXiv 2016

Show all 54 references
  1. [9]

    D.; Monfort, M.; Muller, U.; Zhang, J.; et al

    Bojarski, M.; Del Testa, D.; Dworakowski, D.; Firner, B.; Flepp, B.; Goyal, P.; Jackel, L. D.; Monfort, M.; Muller, U.; Zhang, J.; et al. 2016. End to end learning for self-driving cars. arXiv preprint arXiv:1604.07316

    2016 arXiv

  2. [10]

    o nn, H.; and T \

    Borg, M.; Englund, C.; Wnuk, K.; Duran, B.; Levandowski, C.; Gao, S.; Tan, Y.; Kaijser, H.; L \"o nn, H.; and T \"o rnqvist, J. 2018. Safely entering the deep: A review of verification and validation for machine learning and a challenge elicitation in the automotive industry. ...

    2018 arXiv

  3. [11]

    Clancey, W. J. 1979. Transfer of Rule-Based Expertise through a Tutorial Dialogue . Ph.D. diss., Dept.\ of Computer Science, Stanford Univ., Stanford, Calif

    1979

  4. [12]

    Clancey, W. J. 1983. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education . In Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83) , 556--560. Menlo Park, Calif: IJCAI ...

    1983

  5. [13]

    Clancey, W. J. 1984. Classification Problem Solving . In Proceedings of the Fourth National Conference on Artificial Intelligence, 45--54. Menlo Park, Calif.: AAAI Press

    1984

  6. [14]

    Clancey, W. J. 2021. The Engineering of Qualitative Models . Forthcoming

    2021

  7. [15]

    Deb, K.; Pratap, A.; Agarwal, S.; and Meyarivan, T. 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE transactions on evolutionary computation, 6(2): 182--197

    2002

  8. [16]

    Ehrgott, M.; Ide, J.; and Sch \"o bel, A. 2014. Minmax robustness for multi-objective optimization problems. European Journal of Operational Research, 239(1): 17--31

    2014

  9. [17]

    Engelmore, R.; and Morgan, A., eds. 1986. Blackboard Systems. Reading, Mass.: Addison-Wesley

    1986

  10. [18]

    M.; Fleming, P

    Fonseca, C. M.; Fleming, P. J.; et al. 1993. Genetic algorithms for multiobjective optimization: formulationdiscussion and generalization. In Icga, volume 93, 416--423. Citeseer

    1993

  11. [19]

    April 1, 2024

    Funicello-Paul, L. April 1, 2024. Navigant Research Names Waymo, Ford Autonomous Vehicles, Cruise, and Baidu the Leading Developers of Automated Driving Systems

    2024

  12. [20]

    Goodfellow, I.; Papernot, N.; Huang, S.; Duan, Y.; Abbeel, P.; and Clark, J. 2017. Attacking machine learning with adversarial examples. OpenAI Blog, 24: 1

    2017

  13. [21]

    T.; Liu, Y.; and Chen, Z

    Guo, A.; Zhou, Y.; Tian, H.; Fang, C.; Sun, Y.; Sun, W.; Gao, X.; Luu, A. T.; Liu, Y.; and Chen, Z. 2024. Sovar: Build generalizable scenarios from accident reports for autonomous driving testing. In Proceedings of the 39th IEEE/ACM International Conference on Automated Softwa...

    2024

  14. [22]

    W.; Clancey, W

    Hasling, D. W.; Clancey, W. J.; and Rennels, G. 1984. Strategic explanations for a diagnostic consultation system. International Journal of Man-Machine Studies, 20(1): 3--19

    1984

  15. [23]

    W.; Clancey, W

    Hasling, D. W.; Clancey, W. J.; Rennels, G. R.; and Test, T. 1983. Strategic Explanations in Consultation---Duplicate . The International Journal of Man-Machine Studies, 20(1): 3--19

    1983

  16. [24]

    Hassija, V.; Chamola, V.; Mahapatra, A.; Singal, A.; Goel, D.; Huang, K.; Scardapane, S.; Spinelli, I.; Mahmud, M.; and Hussain, A. 2024. Interpreting black-box models: a review on explainable artificial intelligence. Cognitive Computation, 16(1): 45--74

    2024

  17. [25]

    April 1, 2024

    Hersey, F. April 1, 2024. Baidu launches their open platform for autonomous cars–and we got to test it

    2024

  18. [26]

    Huai, Y. 2023. SORA-SVL Simulator

    2023

  19. [27]

    Ishibuchi, H.; Tsukamoto, N.; and Nojima, Y. 2008. Evolutionary many-objective optimization: A short review. In 2008 IEEE congress on evolutionary computation (IEEE world congress on computational intelligence), 2419--2426. IEEE

    2008

  20. [28]

    Ji, J.; Wang, H.; Wu, C.; Ma, Y.; Sun, X.; and Ji, R. 2024. JM3D & JM3D-LLM: Elevating 3D Representation with Joint Multi-modal Cues. IEEE TPAMI

    2024

  21. [29]

    Kl \"u ck, F.; Zimmermann, M.; Wotawa, F.; and Nica, M. 2019. Genetic algorithm-based test parameter optimization for ADAS system testing. In 2019 IEEE 19th International Conference on Software Quality, Reliability and Security (QRS), 418--425. IEEE. Doi: 10.1109/QRS.2019.00058

    2019

  22. [30]

    Y.; Liu, C.; and Chen, G

    Li, C.; Gao, D. Y.; Liu, C.; and Chen, G. 2014. Impulsive control for synchronizing delayed discrete complex networks with switching topology. Neural Computing and Applications, 24: 59--68

    2014

  23. [31]

    Li, C.; Li, C.; Liao, X.; and Huang, T. 2011. Impulsive effects on stability of high-order BAM neural networks with time delays. Neurocomputing, 74(10): 1541--1550

    2011

  24. [32]

    Li, C.; Yu, W.; and Huang, T. 2014. Impulsive synchronization schemes of stochastic complex networks with switching topology: average time approach. Neural Networks, 54: 85--94

    2014

  25. [33]

    Long, Q. 2014. A constraint handling technique for constrained multi-objective genetic algorithm. Swarm and Evolutionary Computation, 15: 66--79

    2014

  26. [34]

    Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; Agarwal, S.; et al. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 1

    2020 arXiv

  27. [35]

    B.; Mishra, R.; Acharya, A

    Mishra, D. B.; Mishra, R.; Acharya, A. A.; and Das, K. N. 2019. Test case optimization and prioritization based on multi-objective genetic algorithm. In Harmony Search and Nature Inspired Optimization Algorithms: Theory and Applications, ICHSA 2018, 371--381. Springer

    2019

  28. [36]

    NASA . 2015. Pluto: The 'Other' Red Planet. https://www.nasa.gov/nh/pluto-the-other-red-planet. Accessed: 2018-12-06

    2015

  29. [37]

    Onieva, E.; Hern \'a ndez-Jayo, U.; Osaba, E.; Perallos, A.; and Zhang, X. 2015. A multi-objective evolutionary algorithm for the tuning of fuzzy rule bases for uncoordinated intersections in autonomous driving. Information Sciences, 321: 14--30. Doi: 10.1016/j.ins.2015.05.036

    2015 doi

  30. [38]

    Reynolds, L.; and McDonell, K. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended abstracts of the 2021 CHI conference on human factors in computing systems, 1--7

    2021

  31. [39]

    Rice, J. 1986. Poligon: A System for Parallel Problem Solving . Technical Report KSL-86-19, Dept.\ of Computer Science, Stanford Univ

    1986

  32. [40]

    Robinson, A. L. 1980 a . New Ways to Make Microcircuits Smaller. Science, 208(4447): 1019--1022

    1980

  33. [41]

    Robinson, A. L. 1980 b . New Ways to Make Microcircuits Smaller---Duplicate Entry . Science, 208: 1019--1026

    1980

  34. [42]

    P.; Dupont, E.; Ruiz, F

    Romera-Paredes, B.; Barekatain, M.; Novikov, A.; Balog, M.; Kumar, M. P.; Dupont, E.; Ruiz, F. J.; Ellenberg, J. S.; Wang, P.; Fawzi, O.; et al. 2024. Mathematical discoveries from program search with large language models. Nature, 625(7995): 468--475

    2024

  35. [43]

    I.; and Tselikas, N

    Roumeliotis, K. I.; and Tselikas, N. D. 2023. Chatgpt and open-ai models: A preliminary review. Future Internet, 15(6): 192

    2023

  36. [44]

    Tian, H.; Jiang, Y.; Wu, G.; Yan, J.; Wei, J.; Chen, W.; Li, S.; and Ye, D. 2022 a . MOSAT: finding safety violations of autonomous driving systems using multi-objective genetic algorithm. In Proceedings of the 30th ACM Joint European Software Engineering Conference and Sympos...

    2022

  37. [45]

    Tian, H.; Wu, G.; Yan, J.; Jiang, Y.; Wei, J.; Chen, W.; Li, S.; and Ye, D. 2022 b . Generating critical test scenarios for autonomous driving systems via influential behavior patterns. In Proceedings of the 37th IEEE/ACM International Conference on Automated Software Engineer...

    2022

  38. [46]

    Varshney, K. R. 2016. Engineering safety in machine learning. In 2016 Information Theory and Applications Workshop (ITA), 1--5. IEEE

    2016

  39. [47]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention Is All You Need. arXiv:1706.03762

    2017 arXiv

  40. [48]

    Wirsansky, E. 2020. Hands-on genetic algorithms with Python: applying genetic algorithms to solve real-world deep learning and artificial intelligence problems. Packt Publishing Ltd

    2020

  41. [49]

    V.; et al

    Zhang, S.; Roller, S.; Goyal, N.; Artetxe, M.; Chen, M.; Chen, S.; Dewan, C.; Diab, M.; Li, X.; Lin, X. V.; et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

    2022 arXiv

  42. [50]

    N.; and Zhang, Q

    Zhou, A.; Qu, B.-Y.; Li, H.; Zhao, S.-Z.; Suganthan, P. N.; and Zhang, Q. 2011. Multiobjective evolutionary algorithms: A survey of the state of the art. Swarm and evolutionary computation, 1(1): 32--49

    2011

  43. [51]

    Zitzler, E.; Laumanns, M.; and Thiele, L. 2001. SPEA2: Improving the strength Pareto evolutionary algorithm. TIK report, 103

    2001

  44. [52]

    Zitzler, E.; and Thiele, L. 1999. Multiobjective evolutionary algorithms: a comparative case study and the strength Pareto approach. IEEE transactions on Evolutionary Computation, 3(4): 257--271

    1999

  45. [53]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  46. [54]

    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 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.