Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

An LLM that treats text attributes as inheritable genes generates synthetic data that outperforms prior prompting methods in their experiments.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Genetic Prompt uses LLMs to run semantic-level crossover and mutation on text attributes from two far-apart parent examples, producing synthetic data that improves downstream NLP performance, especially for rare classes.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection The premise is promising, but the appendix prompt-template mismatches make the headline results unverifiable as printed. the 4 major comments →

arxiv 2509.02040 v1 pith:DUI3KUPB submitted 2025-09-02 cs.CL

Attributes as Textual Genes: Leveraging LLMs as Genetic Algorithm Simulators for Conditional Synthetic Data Generation

Guangzeng Han , Weisi Liu , Xiaolei Huang This is my paper
classification cs.CL
keywords synthetic data generationlarge language modelsgenetic algorithmscrossover and mutationtext attributesactive learningdata augmentationNLP
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

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 proposes Genetic Prompt, a way to make LLMs generate synthetic text by simulating a genetic algorithm over semantic text attributes. Instead of fixing attribute values ahead of time, it asks the LLM to name the important attributes ('genes'), then chooses two real examples that are far apart in meaning and prompts the LLM to create offspring that inherit some genes from each parent and mutate the rest. The claim is that this produces synthetic data that is both more diverse and closer to the real data distribution, so that downstream models trained on it perform better than with earlier attribute-conditioned or curation-based generation methods. The strongest evidence is in gold-synthetic fusion experiments, where adding Genetic Prompt data raises macro-F1 most on class-imbalanced datasets such as ChemProt, DDI, and StackExchange.

Core claim

The central claim is that an LLM can act as a genetic-algorithm simulator when the units of inheritance are semantic text attributes rather than words or sentences. With two parent examples and a small set of task-specific genes (such as sentence structure, voice, entity proximity, and interaction verb), the LLM performs crossover by assigning some genes to be inherited from one parent and some from the other, and mutation by requiring the remaining genes to differ from both parents while keeping the class or relation intact. The paper argues that this attribute-level evolutionary loop yields offspring with high intra-class diversity and a distribution closer to the gold data, and that this

What carries the argument

Textual genes: a small, task-specific set of semantic and linguistic attributes (e.g., length, voice, entity proximity, interaction verb, readability) treated as the heritable units of text. The LLM is prompted to perform crossover by randomly partitioning these genes between the two parents and mutation by altering the non-inherited genes; parent pairs are selected by an active-learning rule that picks the previously unused pair with the largest embedding distance, enlarging the offspring search space. This machinery is what turns simple prompt-tuning into an evolutionary search over text traits.

Load-bearing premise

The method assumes the LLM reliably obeys the gene-inheritance instructions: when told to take sentence structure from one parent and entity proximity from the other, it must preserve the target relation or class while actually varying the specified attributes-if it ignores the instructions, copies a parent, or drifts off-label, the diversity gains and downstream improvements collapse.

What would settle it

Generate two large synthetic batches with the same parents and the same prompts, but in one batch replace the gene-inheritance instructions with a simple 'write a sentence about this relation' instruction (parents still shown as examples). If downstream F1 is the same for both batches, the genetic mechanism itself is not responsible for the gains. A second check: measure whether the offspring actually differ on the genes the prompt says to mutate; if the assigned genes are unchanged in most outputs, the mechanism is not operating as claimed.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Fusing Genetic Prompt synthetic data with the original training set improves downstream performance on most of the eight tested datasets, with the largest gains on imbalanced sets: +3.2 macro-F1 on ChemProt, +3.1 on DDI, +3.7 on StackExchange, and +4.1 micro-F1 on Conll04.
  • The method is more stable than the baselines across generator sizes: as the generating LLM grows from 3B to 70B, Genetic Prompt's downstream scores keep improving or stay competitive, while SimPrompt, AttrPrompt, and Curated LLM sometimes degrade.
  • Genetic Prompt scales with the amount of synthetic data: downstream performance keeps rising as more offspring are generated, whereas the baselines plateau or decline on ChemProt past about 1-2k samples.
  • Semantic-level genes carry the effect: replacing the gene-based crossover and mutation with word-level operations ('Word as Gene') sharply reduces downstream performance on almost every dataset in the ablation.
  • Even a 14B open-weights generator (Phi-4) under Genetic Prompt can match or beat baseline methods that use GPT-4o on several datasets, suggesting the recipe matters more than raw generator strength.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The active-learning parent selection is effectively farthest-point sampling in embedding space; pairing it with model uncertainty or classifier feedback could push gains further, but the paper does not test that variant.
  • Because the gene set is first elicited from an LLM and then hand-curated per dataset, fully automating gene discovery is a natural next test and would show whether the method transfers to new tasks without human tuning.
  • The fusion gains concentrate in small, domain-specific, and imbalanced datasets; in large balanced settings such as AGNews the benefit is small, so the practical value likely lies in low-resource biomedical and technical domains.
  • Treating attributes as genes is language-agnostic in principle, but the paper only evaluates English; running the same prompts in another language would reveal whether LLMs can execute the genetic operators across different linguistic structures.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Genetic Prompt, a framework in which an LLM is prompted to simulate genetic crossover and mutation over high-level textual attributes ("genes") extracted from seed data, with parent pairs chosen by maximum embedding distance via an active-learning-style selection. The method is evaluated for conditional synthetic data generation on eight datasets spanning classification, relation extraction, and summarization, using several generator LLMs (Phi4, Llama3.1-70B, GPT-3.5, GPT-4o) and downstream models (RoBERTa-base, T5-large). The authors report intrinsic diversity/distribution metrics, downstream F1/Rouge, gold-synthetic fusion gains, ablations, and scaling studies.

Significance. If the reported results hold, the work would offer a practical and modular recipe for conditional synthetic data generation: an LLM-based genetic operator on semantic attributes, avoiding word-level augmentation and per-sample fitness evaluation. The paper includes a code link, multi-dataset and multi-generator evaluation, ablations, and fusion experiments, which are useful. However, the current manuscript contains reproducibility-blocking inconsistencies in the printed prompt templates and several claims that are contradicted by the paper's own tables. These issues must be corrected before the empirical contribution can be assessed.

major comments (4)
  1. [Appendix B.6 (B.6.1–B.6.3)] The prompt templates as printed cannot reproduce the reported experiments. B.6.1 is labeled AGNews but asks for chemical–protein relation sentences with XML tags; B.6.2 is labeled StackExchange but asks for a news article; B.6.3 is labeled ChemProt but asks for a Stack Exchange question. If these templates were actually used, the AGNews, StackExchange, and ChemProt synthetic data would be off-task, making the GPT-4o results in Table 3 (86.7, 49.8, 81.6) and fusion gains in Table 4 implausible. If the experiments used different prompts, the appendix must be corrected and the exact prompts archived. This issue is load-bearing because those three datasets appear in the headline 'consistently outperforms' claim and in the scaling analyses (Figures 2–4).
  2. [Section 4.1, Table 2] The text claims Genetic Prompt "stands out by achieving the best performance in APS, CMD, and vocabulary size." Table 2 contradicts this on two datasets. For StackExchange, AttrPrompt has lower APS (0.306) than Ours (0.324), and Curated LLM has lower CMD (0.520) than Ours (0.595). For MeQSum, AttrPrompt has lower APS (0.296) and lower CMD (0.768) than Ours (0.316 and 0.834). The claim should be restricted to datasets where the table actually supports it, or the evaluation/claim must be revised.
  3. [Section 4.2, Table 3] The abstract and Section 4.2 say Genetic Prompt "significantly outperforms" or "consistently outperforms" the baselines. The table contains multiple cells where a baseline is better. Examples: StackExchange GPT-4o — SimPrompt 52.2 vs. Ours 49.8; DDI Phi4 — AttrPrompt 68.3 vs. Ours 62.2; Conll04 GPT-3.5 — SimPrompt 76.9 and Curated LLM 73.3 vs. Ours 67.5. The claim of consistent superiority across generators is therefore not supported by the reported numbers. The authors should either soften the claim to dataset-model-specific advantages or perform a formal significance/aggregation analysis.
  4. [Sections 2.2–2.3, Appendix C.1] The intrinsic diversity advantage is partly built into the method. Parent selection maximizes the Euclidean distance between sentence embeddings (Section 2.2), while APS measures average cosine similarity of embeddings (Appendix C.1). Selecting maximally distant parents will tend to reduce APS in the offspring even before the crossover/mutation prompt has any effect. The paper does not provide an intrinsic evaluation of the non-selection components: e.g., random parent selection with the same crossover/mutation, or an intrinsic comparison of offspring versus parents. This is important because the diversity claim is used to explain downstream gains.
minor comments (5)
  1. [Algorithm 1] The input includes a diversity coefficient α and computes APS of gold classes, but α is never used in the algorithm or defined in the main text. Please clarify its role or remove it.
  2. [Section 3.2] "temperature and p-value of all LLMs to 1" should read "top-p" rather than "p-value." Also, the exact sampling parameters (temperature, top-p) for each model should be listed, as 'temperature=1' for all models is unusual and may affect reproducibility.
  3. [Table 10, Appendix B.5] The text says "Across all four datasets" but Table 10 reports genes for eight datasets. Please correct the count and the surrounding sentence.
  4. [Table 3] Curated LLM has no entries for SciTLDR and MeQSum. The paper should state why, since Table 2 also lacks these rows; otherwise readers cannot compare fusion gains for these two tasks.
  5. [General] All downstream comparisons report standard deviations over three runs, but no significance tests or confidence intervals are given. Given the small differences in several cells, a paired significance test would help support the 'outperforms' language.

Circularity Check

1 steps flagged

Mild intrinsic-evaluation loop only; downstream claims are external and not circular. Appendix B.6 prompt-template mismatches are a non-circular reproducibility flaw.

specific steps
  1. other [Section 2.2 (Active Parents Selection) vs Section 4.1/Table 2 (Intrinsic Analysis)]
    "The extracted pair of samples is the one that exhibits the largest semantic distance... To quantify this, we compute the Euclidean distance between their semantic representations and choose the pair with the greatest distance. ... APS measures the overall similarity between samples by averaging the cosine similarity between all embeddings."

    The intrinsic diversity result is not fully independent of the method's selection objective. Parents are chosen by maximizing Euclidean distance in Sentence-BERT embedding space, while APS is the mean cosine similarity in that same embedding space. Thus the lower APS reported for Genetic Prompt in Table 2 is partly a direct consequence of the parent-selection rule rather than an emergent property of the generated data. This is only a partial loop: downstream F1/Rouge evaluations and gold-synthetic fusion gains are external benchmarks and do not share this selection objective.

full rationale

The paper is empirical and makes no first-principles derivation, so the circularity burden is low. The one mild loop is in the intrinsic evaluation: parent selection explicitly maximizes pairwise embedding distance, and the APS metric measures mean pairwise embedding similarity in the same Sentence-BERT space, so the reported APS advantage is partly objective-aligned rather than a fully independent finding. This does not affect the central extrinsic claims, which are evaluated on held-out gold test sets and fusion with original training data. Self-citations (Han et al. 2024, Liu et al. 2025, Sharif et al. 2025, Zhang et al. 2025a) appear only in related work or acknowledgments and are not load-bearing. No uniqueness theorem or ansatz is imported via self-citation, and the gene/attribute relabeling is accompanied by concrete crossover/mutation and selection mechanisms. A non-circularity reproducibility flaw should be noted: Appendix B.6 assigns mismatched prompt templates (AGNews template is for ChemProt, StackExchange template is for AGNews, ChemProt template is for StackExchange), which would prevent reproducing the reported results for those datasets if taken literally; this is a correctness/reporting issue, not circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 1 invented entities

The central claim rests on the LLM's ability to perform semantic crossover and mutation correctly, on human-selected per-dataset genes, and on embedding similarity as a proxy for diversity. No new physical entities are introduced; 'textual genes' is a relabeling of attributes. The largest hidden costs are the per-dataset gene sets, the random gene partition, and the target data sizes.

free parameters (5)
  • Per-dataset gene sets = Table 10 (e.g., ChemProt: length, voice, sentence structure, interaction verb, modifier, negation, entity proximity)
    Selected through human-AI collaboration and empirically selected per dataset (Section 2.1); the choice of which attributes are genes defines the search space for crossover and mutation.
  • Initial population size = 50 per class
    Algorithm 1 sets P = {p0,...,p50}; the size is not swept or justified.
  • Random gene partition G1/G2/G3 = random split
    Section 2.3 randomly partitions textual genes into three groups; the random seed is not reported, so the exact crossover assignment is a hidden choice.
  • Target synthetic data size per dataset = AGNews 6k, StackExchange 27k, others 3k (Table 11)
    Sizes are set per dataset; Table 9 shows performance varies with size, so these choices are consequential.
  • Diversity coefficient alpha = unused
    Algorithm 1 declares alpha as an input but the body never references it, making it an unused declared parameter.
axioms (5)
  • domain assumption LLMs faithfully execute crossover and mutation instructions at the semantic attribute level.
    Section 2.3 formulates D_syn from LLM generation; if the LLM does not follow the gene inheritance constraints, the method reduces to paraphrasing.
  • domain assumption Human-selected textual genes cover the relevant axes of variation for each dataset.
    Table 10 lists 4-7 genes per dataset; no coverage check or criterion for gene sufficiency is reported.
  • domain assumption Sentence-BERT embeddings capture semantic diversity and distribution shift in a way that predicts downstream utility.
    Both parent selection (Section 2.2) and intrinsic metrics (Section 4.1, Appendix C) rely on these embeddings.
  • domain assumption Samples generated for class l preserve the gold label l.
    The prompts ask for a sentence about a given relation or category; no label-noise check or human verification on the generated population is reported.
  • domain assumption Baselines are implemented and tuned as intended by their authors and are equally matched.
    Section 3.3 re-implements AttrPrompt, SimPrompt, and CLLM; there is no shared codebase or hyperparameter equivalence analysis.
invented entities (1)
  • Textual genes (semantic attributes treated as alleles) no independent evidence
    purpose: Framing device that lets the LLM perform crossover and mutation over text attributes instead of tokens
    No independent falsifiable handle; it is a relabeling of attributes such as length, voice, and entity proximity. The method's utility is only testable through downstream task performance.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Attributes as Textual Genes: Leveraging LLMs as Genetic Algorithm Simulators for Conditional Synthetic Data Generation." pith.science (2026). https://pith.science/paper/DUI3KUPB

@misc{pith2026250902040,
  author       = {Pith},
  title        = {Pith review of: Attributes as Textual Genes: Leveraging LLMs as Genetic Algorithm Simulators for Conditional Synthetic Data Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DUI3KUPB}},
  note         = {Machine review of arXiv:2509.02040}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Large Language Models (LLMs) excel at generating synthetic data, but ensuring its quality and diversity remains challenging. We propose Genetic Prompt, a novel framework that combines genetic algorithms with LLMs to augment synthetic data generation. Our approach treats semantic text attributes as gene sequences and leverages the LLM to simulate crossover and mutation operations. This genetic process enhances data quality and diversity by creating novel attribute combinations, yielding synthetic distributions closer to real-world data. To optimize parent selection, we also integrate an active learning scheme that expands the offspring search space. Our experiments on multiple NLP tasks reveal several key findings: Genetic Prompt not only significantly outperforms state-of-the-art baselines but also shows robust performance across various generator model sizes and scales. Moreover, we demonstrate that fusing our synthetic data with the original training set significantly boosts downstream model performance, particularly for class-imbalanced scenarios. Our findings validate that Genetic Prompt is an effective method for producing high-quality synthetic data for a wide range of NLP applications.

Figures

Figures reproduced from arXiv: 2509.02040 by Guangzeng Han, Weisi Liu, Xiaolei Huang.

Figure 1
Figure 1. Figure 1: Overview of our Genetic Prompt framework. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effects analysis of generator size and synthetic data scale across diverse NLP tasks. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Empirical analysis on generator size [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Reverse-Engineered Reasoning for Open-Ended Generation

    cs.AI 2025-09 conditional novelty 6.0

    Given a high-quality output, the authors search for a thinking trace that minimizes that output's perplexity, then fine-tune Qwen3-8B on 20,000 such traces, reporting writing performance near GPT-4o and Claude 3.5.

Reference graph

Works this paper leans on

35 extracted references · 33 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    The sentence must discuss about the chemical and protein with the relation { label_def }

    work page

  2. [4]

    In Proceedings of the 2022 Con- ference on Empirical Methods in Natural Language Processing, pages 11653–11669, Abu Dhabi, United Arab Emirates

    ZeroGen: Efficient zero-shot learning via dataset generation. In Proceedings of the 2022 Con- ference on Empirical Methods in Natural Language Processing, pages 11653–11669, Abu Dhabi, United Arab Emirates. Association for Computational Lin- guistics. Dian Yu, Zhou Yu, and Kenji Sagae. 2021. Attribute alignment: Controlling text generation from pre- train...

    work page 2022

  3. [5]

    Not All Features Deserve Attention: Graph-Guided Dependency Learning for Tabular Data Generation with Language Models

    Central moment discrepancy (cmd) for domain-invariant representation learning. In Interna- tional Conference on Learning Representations. Chenhao Zhang, Xi Feng, Yuelin Bai, Xeron Du, Jin- chang Hou, Kaixin Deng, Guangzeng Han, Qinrui Li, Bingli Wang, Jiaheng Liu, Xingwei Qu, Yifei Zhang, Qixuan Zhao, Yiming Liang, Ziqiang Liu, Feiteng Fang, Min Yang, Wen...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  4. [8]

    Relation : { class_name } Text : B.6.2 StackExchange You need to generate synthetic data for the News Classification task

    The ’{ Gene [6]} ’ of the sentence and entities must be different from the given 2 examples . Relation : { class_name } Text : B.6.2 StackExchange You need to generate synthetic data for the News Classification task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write a news article about ’{ class_name } ’ category . The article should fo...

    work page

  5. [17]

    The article must be a news about { label_def }

    work page

  6. [21]

    Category : { class_name } Text : B.6.3 Chemprot You need to generate synthetic data for the Stack Exchange question classification task

    The ’{ Gene [3]} ’ of the sentence and entities must be different from the given 2 examples . Category : { class_name } Text : B.6.3 Chemprot You need to generate synthetic data for the Stack Exchange question classification task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write a Stack Exchange question about ’{ class_name } ’. The ’{...

    work page

  7. [25]

    Category : { class_name } Text : B.6.4 DDI You need to generate synthetic data for Drug - Drug Interaction extraction task

    The ’{ Gene [3]} ’ of the sentence and entities must be different from the given 2 examples . Category : { class_name } Text : B.6.4 DDI You need to generate synthetic data for Drug - Drug Interaction extraction task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write a sentence about ’{ class_name } ’ relation between drug and drug . Th...

    work page

  8. [26]

    The sentence must discuss about the drug and drug with the relation { label_def }

    work page

  9. [32]

    The ’{ Gene [5]} ’ of the sentence should inherit from Example2

    work page

  10. [33]

    Relation : { class_name } Text : B.6.5 Semeval2010 You need to generate synthetic data for Relation Classification task

    The ’{ Gene [6]} ’ of the sentence and entities must be different from the given 2 examples . Relation : { class_name } Text : B.6.5 Semeval2010 You need to generate synthetic data for Relation Classification task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write a sentence about ’{ class_name } ’ relation between 2 entities . The ’{ c...

    work page

  11. [39]

    Relation : { class_name } Text : B.6.6 Conll04 You need to generate synthetic data for Relation Classification task

    The ’{ Gene [5]} ’ of the sentence and entities must be different from the given 2 examples . Relation : { class_name } Text : B.6.6 Conll04 You need to generate synthetic data for Relation Classification task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write a sentence about ’{ class_name } ’ relation between 2 entities . The ’{ class...

    work page

  12. [43]

    The ’{ Gene [3]} ’ of the sentence should inherit from Example2

    work page

  13. [44]

    The ’{ Gene [4]} ’ of the sentence should inherit from Example1

    work page

  14. [45]

    Relation : { class_name } Text : B.6.7 SciTLDR You need to generate synthetic data for the computer science paper abstract summarization task

    The ’{ Gene [5]} ’ of the sentence and entities must be different from the given 2 examples . Relation : { class_name } Text : B.6.7 SciTLDR You need to generate synthetic data for the computer science paper abstract summarization task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write computer science paper abstract and it ’ s summary ...

    work page

  15. [49]

    Output : B.6.8 MeQSum You need to generate synthetic data for the medical question summarization task

    The ’{ Gene [3]} ’ of the sentence and entities must be different from the given 2 examples . Output : B.6.8 MeQSum You need to generate synthetic data for the medical question summarization task . Example 1: { sample1 } Example 2: { sample2 } Your task is to write a medical question and it ’ s summary . The abstract and summary should follow the requirem...

    work page

  16. [50]

    The ’{ Gene [0]} ’ of the sentence should inherit from Example1

    work page

  17. [51]

    The ’{ Gene [1]} ’ of the sentence should inherit from Example2

    work page

  18. [52]

    The ’{ Gene [2]} ’ of the sentence should inherit from Example1

    work page

  19. [53]

    The ’{ Gene [3]} ’ of the sentence and entities must be different from the given 2 examples . Output : C Definition of Diversity and Distribution Metrics In this work, we use two metrics to evaluate the diversity and similarity within the datasets: Av- erage Pairwise Similarity (APS) and V ocabulary size. We also use the Central Moment Discrepancy (CMD) (...

    work page 2022

  20. [54]

    cause”, “from

    measures the difference of probability distri- butions on a compact interval that considers higher order central moments. For our computations, we define the compact interval as [−1, 1]N and con- sider the first five central moments. The CMD between two distributions p and q, with corre- sponding samples X = (X1, . . . , Xn) from p and Y = (Y1, . . . , Yn...

    work page

  21. [55]

    The burst has been caused by water hammer pressure

    work page

  22. [56]

    He had chest pains and headaches from mold in the bedrooms

    work page

  23. [57]

    SimPrompt 1

    The heavy rainstorm caused severe flooding in the downtown area, leading to widespread disruptions and evacuations. SimPrompt 1. The rise in global temperatures is causing the melt of polar ice caps. And 2. The intense heatwave led to crop failures in several rural communities, significantly impacting the livelihoods of local farmers. Curated LLM 3. Smoki...

    work page

  24. [58]

    Increase the budget for the marketing team, as the new strategies will cause a significant impact on sales

    work page

  25. [59]

    Is the landscape shaped by erosion?

    work page

  26. [60]

    Genetic Prompt

    The storm was responsible for causing damage, and as a result, the power outage occurred because of it. Genetic Prompt

    work page

  27. [61]

    With the increase in solar activity, satellites experience disruptions in their communications systems

    work page

  28. [62]

    The political upheaval resulted in the region a devastating financial setback due to the scandal surrounding the mismanagement of public funds during the last fiscal cycle

    work page

  29. [63]

    Table 12: Text Examples in Cause-effect class in Semeval dataset

    She observed changes in the plant growth following the application of the fertilizer after the rainfall event. Table 12: Text Examples in Cause-effect class in Semeval dataset

    work page

  30. [2019]

    B.2 Additional Technical Details During generation, for responses that do not adhere to the instructions, such as ‘I am just a large lan- guage model

    a publicly available dataset of consumer health questions from the MedHelp forum paired with expert-written abstractive summaries. B.2 Additional Technical Details During generation, for responses that do not adhere to the instructions, such as ‘I am just a large lan- guage model. ’or ‘I cannot do that for you. ’ , are filtered out. The hyperparameters ca...

    work page 2021

  31. [2022]

    Preprint, arXiv:2206.08896

    Evolution through large models. Preprint, arXiv:2206.08896. Junlong Li, Jinyuan Wang, Zhuosheng Zhang, and Hai Zhao. 2024a. Self-prompting large language models for zero-shot open-domain QA. In Proceedings of the 2024 Conference of the North American Chap- ter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Pa...

    Pith/arXiv arXiv 2024

  32. [2023]

    B.4 Pseudo Code of Genetic Prompt We provide the pseudo code of Genetic Prompt in Algorithm 1

    library to ensure efficient inference. B.4 Pseudo Code of Genetic Prompt We provide the pseudo code of Genetic Prompt in Algorithm 1. B.5 Selected Textual Genes The selection of genes for each dataset was tailored to capture the unique characteristics and require- ments of their respective domains. Table 10 shows the textual genes selected for the differe...

    work page

  33. [2024]

    Doubling Your Data in Minutes: Ultra-fast Tabular Data Generation via LLM-Induced Dependency Graphs

    Large language models as optimizers. In The Twelfth International Conference on Learning Representations. Shuo Yang, Zheyu Zhang, Bardh Prenkaj, and Gjergji Kasneci. 2025. Doubling your data in minutes: Ultra- fast tabular data generation via llm-induced depen- dency graphs. Preprint, arXiv:2507.19334. Jiacheng Ye, Jiahui Gao, Qintong Li, Hang Xu, Jiangta...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  34. [2025]

    arXiv preprint arXiv:2503.02390

    Reso: A reward-driven self-organizing llm- based multi-agent system for reasoning tasks. arXiv preprint arXiv:2503.02390. A Additional Results A.1 Supplementary Experimental Results Due to space limitations, we present the remaining performance metrics of the three experiments Ex- trinsic Evaluation on NLP Tasks and Ablation Study in Table 6 and Table 7 r...

    Pith/arXiv arXiv 2015

  35. [7817]

    Jiawei Chen, Xinyan Guan, Qianhao Yuan, Guozhao Mo, Weixiang Zhou, Yaojie Lu, Hongyu Lin, Ben He, Le Sun, and Xianpei Han

    Curran Associates, Inc. Jiawei Chen, Xinyan Guan, Qianhao Yuan, Guozhao Mo, Weixiang Zhou, Yaojie Lu, Hongyu Lin, Ben He, Le Sun, and Xianpei Han. 2025. Consistentchat: Building skeleton-guided consistent dialogues for large language models from scratch. arXiv preprint arXiv:2506.03558. John Chung, Ece Kamar, and Saleema Amershi. 2023. Increasing diversit...

    work page arXiv 2025

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.