Automatic Prompt Optimization Techniques: Exploring the Potential for Synthetic Data Generation

Gerrit Meixner; Marius Heitlinger; Nina Freise; Ruben Nuredini

REVIEW 3 major objections 4 minor 31 references

Automatic Prompt Optimization Techniques: Exploring the Potential for Synthetic Data Generation

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Automatic prompt optimization can refine prompts without real data, and combining its method families could unlock privacy-safe synthetic data, a systematic review argues.

desk verdict A transparent PRISMA review whose central 'data-free' premise is contradicted by two of its six included methods; the taxonomy is useful but the healthcare synthetic-data extrapolation needs major rework. read the letter →

arxiv 2502.03078 v2 pith:AB2UVN5J submitted 2025-02-05 cs.HC cs.LG

Nina Freise , Marius Heitlinger , Ruben Nuredini , Gerrit Meixner This is my paper · ORCID

classification cs.HCcs.LG

keywords automaticpromptoptimizationsyntheticdatagenerationdata-freelargelanguagemodelssystematicreviewhealthcareprivacyiterativerefinementengineering

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 is a systematic review asking whether automatic prompt optimization can generate synthetic data without access to real datasets, a question that matters for privacy-restricted fields such as healthcare. The review examines six peer-reviewed studies from 2020 to 2024 and groups their methods into feedback-driven, error-based, control-theoretic, and contextual/evolutionary families. Its central claim is that all of these families can iteratively refine prompts without direct access to real data, and that future work should combine them into an integrated framework to improve synthetic data quality. The payoff would be a path to generate realistic domain-specific data while keeping patient records private.

What carries the argument

The central object is the data-free prompt-optimization loop: generate a prompt, evaluate the model's response, then rewrite the prompt based on that evaluation. The review's inclusion criteria—methods must use automatic iterative text-prompt optimization and must not use a training set or open dataset for evaluation—are what make this loop load-bearing, because they restrict the survey to techniques whose prompt improvements do not come from real data. Each of the six studies instantiates the loop differently: critic feedback aggregation, success/failure analysis, error diagnosis, control-theoretic adjustment, or sentence-level mutation under a contextual bandit. The loop is the mechanism that connects prompt optimization to synthetic data generation, since improving prompts without real data is what would allow generating data without real data.

What would settle it

Run one of the reviewed methods, such as STRAGO, on a clinical-note generation task with no labeled examples, so there are no successful or failed cases to analyze, and compare the output of iterated prompts against a fixed manually written prompt. If iteration produces no measurable gain in output plausibility or downstream utility, the claim that these methods are data-free prompt optimizers is contradicted.

Watch

Extended reading notes

Core claim

On the paper's own terms, automatic prompt optimization is a viable route to data-free synthetic data generation. The review's abstract summarizes the landscape in three families—feedback-driven, error-based, and control-theoretic—and its results section adds a contextual/evolutionary family; in both framings, each family is presented as capable of iterative prompt refinement without direct access to real data. The six reviewed methods iterate on a text prompt without training on a dataset: PACE uses an actor-critic loop with aggregated critic feedback, STRAGO analyzes successful and failed cases to avoid prompt drift, REPROMPT and Automatic Behavior Optimization diagnose recurring errors and refine prompts at specific failure points, the optimal-control formulation adjusts prompts through multi-round structured interaction, and the evolutionary approach mutates individual sentences of long prompts under a contextual bandit. The paper concludes that these families are complementary and that an integrated framework—for example, combining error-focused precision with evolutionary scalability—would reduce manual intervention and improve the realism of generated synthetic data. Because the review deliberately excluded methods that rely on training sets or open datasets, the argument is specifically about data-free operation.

Load-bearing premise

The load-bearing premise is that the six reviewed methods can optimize prompts with zero access to real data or ground-truth labels; if that premise fails, the proposed path to privacy-safe synthetic medical data does not follow.

Editorial extensions

If this is right

If the reviewed methods refine prompts without real data, synthetic medical text can be generated without exposing patient records, lowering the privacy barrier to AI development in healthcare.
An integrated framework that combines feedback-driven, error-based, and evolutionary methods could produce higher-fidelity synthetic data with less manual prompt crafting than any single family.
Because the optimization is automatic and iterative, the technique could be embedded into AI systems as a self-improving feature, reducing the burden on users to articulate intent precisely.
The review's exclusion of methods that use real data means the conclusion applies only to a narrow subset of prompt optimizers; extending it to hybrid approaches would require additional validation.

Reading between the lines

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

Beyond the paper: the data-free property is asserted by inclusion criteria rather than demonstrated by an end-to-end test. STRAGO's analysis of successful and failed cases and Ma et al.'s step-level failure identification appear to require labeled or known-correct responses, so a direct transfer to unlabeled clinical text is an open question.
Beyond the paper: a decisive evaluation would be a shared no-ground-truth benchmark in which synthetic clinical notes generated before and after optimization are scored by downstream utility, such as performance of a classifier trained on the synthetic data and tested on a held-out real dataset.
Beyond the paper: the proposed integrated framework could be built as a modular pipeline where an error-diagnosis module identifies failure modes and an evolutionary module explores prompt variations, but the paper stops at recommending the combination rather than specifying how modules share state.

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.

Desk Editor's Note

read the letter

The paper is a systematic review of six automatic prompt optimization techniques, framed around synthetic data generation without direct access to real data. What it does well: the PRISMA process is transparent, the inclusion/exclusion criteria are explicit, the limitations are candidly acknowledged, and the taxonomy (feedback-driven, error-focused, control-theoretic, and evolutionary/contextual) is a workable way to organize a small literature. The research agenda—combining complementary techniques into an integrated framework—is a sensible direction, even if speculative.

The soft spot is not minor. The review's load-bearing premise, stated in the Introduction and Methods, is that all six methods can operate without any established form of ground truth. That does not survive contact with the paper's own descriptions. STRAGO is described as 'analyzing both successful and failed cases' and 'dissecting both correct and incorrect predictions'—you cannot classify a prediction as correct or incorrect without labels. Ma et al.'s Automatic Behavior Optimization requires the optimizer to 'identify failure steps,' which again presumes a correctness signal. The original papers confirm this. So the 'data-free' category is at least partly empty, and the extrapolation to privacy-restricted healthcare synthetic data is unsupported by the evidence presented.

There is also a smaller internal inconsistency: the abstract and conclusion say 'three approaches,' but the Results section presents four distinct categories (Section 3.4 gives Hsieh et al.'s evolutionary method). That looks like a hasty synthesis and is easily fixed.

On the other points: the decision to search only Google Scholar and restrict to open-access or institutionally available papers is a real limitation, but the authors say so. The small sample size (six studies) limits generality, but for a niche framing that is defensible. The summaries of the individual methods are fair and balanced.

The stress-test note holds up on reading the full text. The central claim is not fraudulent—the authors likely believed their inclusion criteria excluded labeled feedback—but it is mistaken, and the mistake undermines the paper's main application story.

Who gets value from this? A reader new to prompt optimization who wants a quick map of recent method families. It is not a reliable source for the claim that these methods work on private medical data without any real data.

Recommendation: it deserves a serious referee, but only with the expectation of major revision. The taxonomy is worth keeping, but the data-free criterion must be re-examined, the abstract corrected, and the healthcare claims either re-evidenced or reframed as open questions.

Referee Report

3 major / 4 minor

Summary. The paper is a PRISMA-guided systematic review of automatic prompt optimization methods, motivated by the possibility of using LLMs to generate synthetic data in privacy-sensitive domains (e.g., healthcare) without direct access to real datasets. The authors screened 57 records from Google Scholar and retained six studies, which the body organizes into four methodological families: feedback-based (PACE, STRAGO), error-focused (REPROMPT, Ma et al.), control-theoretic (Luo et al.), and contextual/evolutionary (Hsieh et al.). The abstract, however, reports only three approaches. The authors argue that the reviewed techniques demonstrate promising prompt-refinement capabilities and call for an integrated framework combining these families to enhance synthetic data generation.

Significance. The paper's main strengths are procedural: the PRISMA flow is reported, the search string is explicit, and the limitations (single database, open-access and language restrictions, dataset-based exclusion) are candidly listed. As a descriptive map of recent prompt-optimization strategies, the taxonomy is a useful starting point and the transparency of the search process is commendable. However, the central claim that these methods operate without direct access to real datasets is not supported by the paper's own descriptions of the included studies, and the extrapolation to synthetic data generation is not grounded in any of the six selected papers. The review is therefore more reliable as a survey of optimization families than as evidence for data-free synthetic data generation.

major comments (3)

[Methods (Table 1) and Sections 3.1–3.2] The inclusion criteria in Table 1 exclude studies that 'utilize a training set or open dataset for evaluation,' and the Methods state that eligible methods must operate without 'any established form of ground truth.' Yet Section 3.1 describes STRAGO as 'analyzing both successful and failed cases' and 'dissecting both correct and incorrect predictions,' and Section 3.2 describes Ma et al.'s Automatic Behavior Optimization as 'requir[ing] the optimizer to identify failure steps.' Identifying correct versus incorrect predictions and failure steps requires a correctness signal that the paper's own criteria exclude; without such a signal, the feedback loops of these methods cannot be computed. The review therefore does not establish that the six methods are data-free; at most they are low-resource or small-feedback methods. The authors should revise the scope statement, re-examine each included study for its actual data requirements, or provide evidence of fully data-free variants.
[Abstract vs. Sections 3.1–3.4] The abstract reports 'three approaches: feedback-driven, error-based, and control-theoretic,' but the Results present four families: feedback-based (PACE, STRAGO), error-focused (REPROMPT, Ma et al.), control-theoretic (Luo et al.), and contextual/evolutionary (Hsieh et al.). The Discussion (Section 4) also lists four families. This inconsistency should be corrected; if the fourth family is deliberately subsumed under one of the other categories, the abstract should say so explicitly.
[Sections 3 and 4] The stated motivation is synthetic data generation, but none of the six retained studies generates synthetic data or evaluates output authenticity, clinical accuracy, or distributional fidelity; they address general tasks such as planning, reasoning, and instruction following. Consequently, the Conclusion's claims that these methods 'display immense potential for generating synthetic data' and that an integrated framework would 'enhance synthetic data generation' go beyond the evidence. The authors should explicitly frame this as an untested research hypothesis and specify what evaluations (e.g., fidelity metrics, downstream task performance, privacy audits) would be needed to support it.

minor comments (4)

[Introduction] The phrase 'commonly refereed to as prompts' should read 'commonly referred to as prompts,' and 'heterogenous' should be 'heterogeneous.'
[Section 2 / Figure 1] The text reports that 15 papers were excluded after full-text screening but does not list the exclusion reasons; please make the reasons explicit either in the text or in a note to Figure 1.
[Section 3.4] The description of Hsieh et al. says 'without direct feedback from task performance, the method may struggle to adapt,' but the preceding sentences state that each revised prompt is evaluated and successful mutations inform future modifications; clarify how task performance is or is not used in the contextual bandit.
[References] Several references are arXiv identifiers with inconsistent URL formats; please ensure all entries include stable and complete links.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the review's taxonomy is an inductive summary of external studies, not a derivation that reduces to its own inputs.

full rationale

This paper is a PRISMA systematic review; its output is a taxonomy and synthesis of six external studies. It performs no fitting, no parameter estimation, and no derivation whose conclusion is equivalent to an input. The category labels (feedback-driven, error-based, control-theoretic, contextual/evolutionary) are inductive summaries of the reviewed methods, which is the normal operation of a review. There are no self-citations by the authors in the reference list; the load-bearing content is the external papers' reported results, so no self-citation chain forces the conclusion. The inconsistency between the stated 'data-free' inclusion criterion and some reviewed methods' need for correct/incorrect predictions (e.g., STRAGO analyzing successful and failed cases, Ma et al. requiring failure-step identification) is a substantive validity and correctness concern, but it is not circular: the review's conclusion is not defined in terms of that criterion, and the synthesis does not reduce to the inclusion criterion by construction. Accordingly, no circular step can be exhibited with equation-level or fitted-parameter-renamed-as-prediction evidence.

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

The paper introduces no free parameters, fitted numbers, or new entities. Its claims rest on the representativeness and data-free nature of the six selected studies and on the transferability of their results to synthetic data generation, all of which are domain assumptions that the authors state or imply but do not verify.

assumptions (3)

domain assumption The six selected studies are representative of automatic data-free prompt optimization methods.
The review generalizes from six studies to the entire landscape of data-free prompt optimization for synthetic data; this is assumed rather than demonstrated.
domain assumption The reviewed methods can operate without access to real datasets or ground-truth labels.
The inclusion/exclusion criteria (Table 1) assert data-free operation, but several included methods appear to rely on labeled examples for evaluation; the data-free property is asserted, not proven.
domain assumption Performance improvements on general reasoning and instruction tasks transfer to synthetic data generation in privacy-sensitive domains.
The six studies target general task accuracy, not synthetic data quality or authenticity; the review's central application claim rests on this transfer assumption.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Automatic Prompt Optimization Techniques: Exploring the Potential for Synthetic Data Generation." pith.science (2026). https://pith.science/paper/AB2UVN5J

@misc{pith2026250203078,
  author       = {Pith},
  title        = {Pith review of: Automatic Prompt Optimization Techniques: Exploring the Potential for Synthetic Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AB2UVN5J}},
  note         = {Machine review of arXiv:2502.03078}
}

read the original abstract

Artificial Intelligence (AI) advancement is heavily dependent on access to large-scale, high-quality training data. However, in specialized domains such as healthcare, data acquisition faces significant constraints due to privacy regulations, ethical considerations, and limited availability. While synthetic data generation offers a promising solution, conventional approaches typically require substantial real data for training generative models. The emergence of large-scale prompt-based models presents new opportunities for synthetic data generation without direct access to protected data. However, crafting effective prompts for domain-specific data generation remains challenging, and manual prompt engineering proves insufficient for achieving output with sufficient precision and authenticity. We review recent developments in automatic prompt optimization, following PRISMA guidelines. We analyze six peer-reviewed studies published between 2020 and 2024 that focus on automatic data-free prompt optimization methods. Our analysis reveals three approaches: feedback-driven, error-based, and control-theoretic. Although all approaches demonstrate promising capabilities in prompt refinement and adaptation, our findings suggest the need for an integrated framework that combines complementary optimization techniques to enhance synthetic data generation while minimizing manual intervention. We propose future research directions toward developing robust, iterative prompt optimization frameworks capable of improving the quality of synthetic data. This advancement can be particularly crucial for sensitive fields and in specialized domains where data access is restricted, potentially transforming how we approach synthetic data generation for AI development.

Figures

Figures reproduced from arXiv: 2502.03078 by the authors.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 14 canonical work pages

[1]

Ap- plied Sciences13(12) (2023), https://www.mdpi.com/2076-3417/13/12/7082

Aldoseri, A., Al-Khalifa, K.N., Hamouda, A.M.: Re-thinking data strategy and integration for artificial intelligence: Concepts, opportunities, and challenges. Ap- plied Sciences13(12) (2023), https://www.mdpi.com/2076-3417/13/12/7082

work page 2023
[2]

Amatriain, X.: Prompt design and engineering: Introduction and advanced meth- ods (2024), https://arxiv.org/abs/2401.14423

arXiv 2024
[3]

CoRR abs/2005.14165 (2020), https://arxiv

Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Win- ter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford...

arXiv 2020
[4]

Frontiers in Neuroscience17 (2023), https://www.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2023.1219133

Carrle, F.P., Hollenbenders, Y., Reichenbach, A.: Generation of synthetic eeg data for training algorithms supporting the diagnosis of major depressive disorder. Frontiers in Neuroscience17 (2023), https://www.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2023.1219133

arXiv 2023
[5]

Chen, W., Koenig, S., Dilkina, B.: Reprompt: Planning by automatic prompt en- gineering for large language models agents (2024),https://arxiv.org/abs/2406. 11132

work page 2024
[6]

Dong, Y., Luo, K., Jiang, X., Jin, Z., Li, G.: Pace: Improving prompt with actor- critic editing for large language model (2024), https://arxiv.org/abs/2308. 10088

work page 2024
[7]

IEEE Journal of Biomedical and Health Informatics 28(7), 4084–4093 (2024)

Dorjsembe, Z., Pao, H.K., Odonchimed, S., Xiao, F.: Conditional diffusion mod- els for semantic 3d brain mri synthesis. IEEE Journal of Biomedical and Health Informatics 28(7), 4084–4093 (2024)

work page 2024
[8]

Neurocomputing321, 321–331 (Dec 2018)

Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., Greenspan, H.: Gan-based synthetic medical image augmentation for increased cnn performance in liver lesion classification. Neurocomputing321, 321–331 (Dec 2018)

2018

Show all 31 references

[9]

Freise, M

Goodfellow, I.J., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., Bengio, Y.: Generative adversarial networks (2014),https:// arxiv.org/abs/1406.2661 12 N. Freise, M. Heitlinger et al

2014 arXiv
[10]

IEEE Intelligent Systems24(2), 8–12 (2009)

Halevy, A., Norvig, P., Pereira, F.: The unreasonable effectiveness of data. IEEE Intelligent Systems24(2), 8–12 (2009)

2009
[11]

CoRR abs/2006.11239 (2020), https://arxiv.org/abs/2006.11239

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. CoRR abs/2006.11239 (2020), https://arxiv.org/abs/2006.11239

2020 arXiv
[12]

Hsieh, C.J., Si, S., Yu, F.X., Dhillon, I.S.: Automatic engineering of long prompts (2023), https://arxiv.org/abs/2311.10117

2023 arXiv
[13]

Kepel, D., Valogianni, K.: Autonomous prompt engineering in large language mod- els (2024), https://arxiv.org/abs/2407.11000

2024 arXiv
[14]

Kingma, D.P., Welling, M.: Auto-encoding variational bayes (2013), https:// arxiv.org/abs/1312.6114

2013 arXiv
[15]

In: Machine Learning and Knowledge Discovery in Databases

Kumichev, G., Blinov, P., Kuzkina, Y., Goncharov, V., Zubkova, G., Zenovkin, N., Goncharov, A., Savchenko, A.: Medsyn: Llm-based synthetic medical text gener- ation framework. In: Machine Learning and Knowledge Discovery in Databases. Applied Data Science Track: European Confe...

2024
[16]

Annals of internal medicine151(4), W–65 (2009)

Liberati, A., Altman, D.G., Tetzlaff, J., Mulrow, C., Gøtzsche, P.C., Ioannidis, J.P.,Clarke,M.,Devereaux,P.J.,Kleijnen,J.,Moher,D.:Theprismastatementfor reporting systematic reviews and meta-analyses of studies that evaluate health care interventions: explanation and elaborat...

2009
[17]

Liu, C., Zhang, W., Zhao, Y., Luu, A.T., Bing, L.: Is translation all you need? a study on solving multilingual tasks with large language models (2024),https: //arxiv.org/abs/2403.10258

2024 arXiv
[18]

Luo, Y., Tang, Y., Shen, C., Zhou, Z., Dong, B.: Prompt engineering through the lens of optimal control (2023),https://arxiv.org/abs/2310.14201

2023 arXiv
[19]

Ma, R., Wang, X., Zhou, X., Li, J., Du, N., Gui, T., Zhang, Q., Huang, X.: Are large language models good prompt optimizers? (2024),https://arxiv.org/abs/ 2402.02101

2024 arXiv
[20]

Minaee, S., Mikolov, T., Nikzad, N., Chenaghlu, M., Socher, R., Amatriain, X., Gao, J.: Large language models: A survey (2024),https://arxiv.org/abs/2402. 06196

2024
[21]

OpenAI (2023), https://openai.com/index/ gpt-4/

OpenAI: Gpt-4 technical report. OpenAI (2023), https://openai.com/index/ gpt-4/

2023
[22]

Rezende, D.J., Mohamed, S., Wierstra, D.: Stochastic backpropagation and ap- proximate inference in deep generative models (2014),https://arxiv.org/abs/ 1401.4082

2014 arXiv
[23]

Schulhoff, S., Ilie, M., Balepur, N., Kahadze, K., Liu, A., Si, C., Li, Y., Gupta, A., Han, H., Schulhoff, S., Dulepet, P.S., Vidyadhara, S., Ki, D., Agrawal, S., Pham, C., Kroiz, G., Li, F., Tao, H., Srivastava, A., Costa, H.D., Gupta, S., Rogers, M.L., Goncearenco, I., Sarli...

2024 arXiv
[24]

JMIR Hum Factors11, e47031 (Jan 2024), https://humanfactors.jmir.org/2024/1/e47031

Shevtsova, D., Ahmed, A., Boot, I.W.A., Sanges, C., Hudecek, M., Jacobs, J.J.L., Hort, S., Vrijhoef, H.J.M.: Trust in and acceptance of artificial intelligence appli- cations in medicine: Mixed methods study. JMIR Hum Factors11, e47031 (Jan 2024), https://humanfactors.jmir.org...

2024
[25]

Information15(5) (2024), https://www.mdpi.com/2078-2489/15/5/264 Automatic Prompt Optimization Techniques 13

Sufi, F.: Addressing data scarcity in the medical domain: A gpt-based approach for synthetic data generation and feature extraction. Information15(5) (2024), https://www.mdpi.com/2078-2489/15/5/264 Automatic Prompt Optimization Techniques 13

2024
[26]

http://www.incompleteideas.net/IncIdeas/ (2019), accessed: December 2024

Sutton, R.: The bitter lesson. http://www.incompleteideas.net/IncIdeas/ (2019), accessed: December 2024

2019
[27]

Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., Rodriguez, A., Joulin, A., Grave, E., Lample, G.: Llama: Open and efficient foundation language models (02 2023)

2023
[28]

Uyguner, I., Martin, I., Pallaoro, P., Boeker, M., Modersohn, L.: Meeting ab- stract. In: 68. Jahrestagung der Deutschen Gesellschaft für Medizinische Infor- matik, Biometrie und Epidemiologie e. V. (GMDS). p. DocAbstr. 260. German Medical Science GMS Publishing House (2023), ...

2023
[29]

Wu, Y., Gao, Y., Zhu, B.B., Zhou, Z., Sun, X., Yang, S., Lou, J.G., Ding, Z., Yang, L.: Strago: Harnessing strategic guidance for prompt optimization (2024), https://arxiv.org/abs/2410.08601

2024 arXiv
[30]

ArXiv abs/2403.17549 (2024), https://api.semanticscholar.org/CorpusID: 268691510

Xu, J., Wu, B., Huang, J., Gong, Y., Zhang, Y., Liu, B.: Practical applications of advanced cloud services and generative ai systems in medical image analysis. ArXiv abs/2403.17549 (2024), https://api.semanticscholar.org/CorpusID: 268691510

2024 arXiv
[31]

arXiv preprint arXiv:2309.03409 (2023)

Yang, C., Wang, X., Lu, Y., Liu, H., Le, Q.V., Zhou, D., Chen, X.: Large language models as optimizers. arXiv preprint arXiv:2309.03409 (2023)

2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [1]

Ap- plied Sciences13(12) (2023), https://www.mdpi.com/2076-3417/13/12/7082

Aldoseri, A., Al-Khalifa, K.N., Hamouda, A.M.: Re-thinking data strategy and integration for artificial intelligence: Concepts, opportunities, and challenges. Ap- plied Sciences13(12) (2023), https://www.mdpi.com/2076-3417/13/12/7082

work page 2023

[2] [2]

Amatriain, X.: Prompt design and engineering: Introduction and advanced meth- ods (2024), https://arxiv.org/abs/2401.14423

arXiv 2024

[3] [3]

CoRR abs/2005.14165 (2020), https://arxiv

Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Win- ter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford...

arXiv 2020

[4] [4]

Frontiers in Neuroscience17 (2023), https://www.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2023.1219133

Carrle, F.P., Hollenbenders, Y., Reichenbach, A.: Generation of synthetic eeg data for training algorithms supporting the diagnosis of major depressive disorder. Frontiers in Neuroscience17 (2023), https://www.frontiersin.org/journals/ neuroscience/articles/10.3389/fnins.2023.1219133

arXiv 2023

[5] [5]

Chen, W., Koenig, S., Dilkina, B.: Reprompt: Planning by automatic prompt en- gineering for large language models agents (2024),https://arxiv.org/abs/2406. 11132

work page 2024

[6] [6]

Dong, Y., Luo, K., Jiang, X., Jin, Z., Li, G.: Pace: Improving prompt with actor- critic editing for large language model (2024), https://arxiv.org/abs/2308. 10088

work page 2024

[7] [7]

IEEE Journal of Biomedical and Health Informatics 28(7), 4084–4093 (2024)

Dorjsembe, Z., Pao, H.K., Odonchimed, S., Xiao, F.: Conditional diffusion mod- els for semantic 3d brain mri synthesis. IEEE Journal of Biomedical and Health Informatics 28(7), 4084–4093 (2024)

work page 2024

[8] [8]

Neurocomputing321, 321–331 (Dec 2018)

Frid-Adar, M., Diamant, I., Klang, E., Amitai, M., Goldberger, J., Greenspan, H.: Gan-based synthetic medical image augmentation for increased cnn performance in liver lesion classification. Neurocomputing321, 321–331 (Dec 2018)

2018