Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Evolving Prompts In-Context: An Open-ended, Self-replicating Perspective

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Pruning ICL demonstrations into gibberish can match or beat automatic prompt optimizers.

desk verdict A real pruning effect, buried under an 'always' claim the paper's own tables contradict. read the letter →

arxiv 2506.17930 v1 pith:4UX77TDU submitted 2025-06-22 cs.AI cs.CLcs.LGcs.NEcs.RO

Jianyu Wang , Zhiqiang Hu , Lidong Bing This is my paper
classification cs.AIcs.CLcs.LGcs.NEcs.RO
keywords in-contextlearningpromptoptimizationevolutionarysearchpruninggibberishpromptspartialcontexthypothesislargelanguagemodelalignment
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that an in-context learning (ICL) prompt can be improved by deleting tokens until what remains looks like meaningless gibberish, and that such pruning can match or exceed the performance of state-of-the-art automatic prompt optimizers. The authors call this the Partial Context Hypothesis: from a natural-language prompt x there is a subsequence z that performs far better, even when the original prompt was weak. They show that a simple hill-climbing pruner already improves classification, style transfer, multiple-choice QA, and math reasoning across several language models, and that their evolutionary framework, PromptQuine, finds pruned prompts that beat competitive instruction-optimization baselines. If the finding holds, prompt engineering need not be about writing clearer instructions; it may be about searching a large space of token deletions, and even the tokens a human would keep can be the wrong ones.

What carries the argument

The carrying object is the Partial Context Hypothesis, coupled with an evolutionary search over binary token masks: each mask (genotype) selects which tokens of the original ICL prompt survive, and the pruned prompt is the phenotype. PromptQuine evolves these masks by copy-then-mutate (bit-flips 1 to 0, i.e., pruning random tokens), with fitness evaluated by the piecewise label-probability reward function (Equation 3), selection via tournament and regularized evolution (only new offspring compete), and a final calibration-then-selection step that re-ranks elite prompts by validation accuracy. The key mechanism is that the search space is the subsequence lattice of the original prompt, so no new tokens are ever introduced; the optimizer explores which absences matter.

What would settle it

Take a fixed model and task, hold out a test set disjoint from the 200-sample validation slice, and run PromptQuine on a fresh random ICL prompt: if the best pruned subsequence does not beat both the original prompt and a matched-length random subsequence of the same prompt on the test set, the Partial Context Hypothesis fails. A direct check is to compare PROMPTQUINE's test-set accuracy to raw 1-shot ICL and to a random-subsequence baseline controlling for length across the six classification datasets.

Watch

Extended reading notes

Core claim

The paper's central claim, on its own terms, is that pruning random demonstrations into seemingly incoherent gibberish reliably improves ICL performance, and that the best pruned prompts always match or surpass state-of-the-art automatic prompt optimization techniques. The discovery is the Partial Context Hypothesis: every natural ICL prompt x has a subsequence z (found by search) whose task performance Z exceeds the original X, and this Z can approach or exceed the performance of token-level search artifacts like the prompts produced by RLPrompt. PromptQuine is the demonstration that this pruning strategy can be discovered automatically in low-data regimes, using only the tokens already present in the context, and that the discovered subsequences transfer to test performance. A notable secondary discovery is that pruning preserves label words in demonstrations and that removing them degrades performance, while pruning can also recover above-chance accuracy even when the verbalizers are random words.

Load-bearing premise

The search assumes that the 200-sample validation score computed with the piecewise label-probability reward is a faithful stand-in for true test performance; if that proxy is misaligned on some tasks, the evolved gibberish may just be overfitting a narrow validation slice.

Editorial extensions

If this is right

  • Pruning a 1-shot ICL prompt with PromptQuine matches or beats EvoPrompt, Promptbreeder, RLPrompt, and the LLMLingua compressors on the six classification datasets reported in Table 1.
  • Because the mutation operator only deletes tokens, the optimized prompts are shorter, so serving the model is faster while performance is higher.
  • The gains persist across model families and alignment levels, including base, SFT, and RLHF models, suggesting the effect is not an artifact of one model's training.
  • Scaling the initial context to 4-shot makes pruning even more effective, so richer contexts unlock additional pruning gains.
  • The same pipeline nearly doubles jailbreak success rate under a simple priming setup, implying that unnatural-language prompts are a real alignment vulnerability.
  • The label-word intervention results show that pruned prompts retain the same sensitivity to label words as conventional ICL, pointing to a shared mechanism rather than a purely adversarial exploit.

Reading between the lines

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

  • Beyond the paper, the same subsequence search could apply to any fixed prompt text, such as system prompts or API-level instructions, not just ICL demonstrations.
  • The template sensitivity PromptQuine exhibits (up to 14.3% accuracy spread on SNLI when only the template changes) suggests the discovered gibberish is not a universal shortcut but a model-template interaction; a testable extension is to measure how much of the pruning gain transfers across templates and across models.
  • The random-verbalizer results hint that pruning helps the model use its label-space prior more efficiently, so combining pruning with label-space calibration methods could yield further gains.
  • Because the search is constrained to deletions, a natural extension is to allow controlled insertions or replacements; the paper's own limitation analysis suggests such mutations could reduce the current instability.
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

3 major / 5 minor

Summary. This paper proposes PROMPTQUINE, an evolutionary search framework that optimizes in-context learning (ICL) prompts by pruning tokens from random demonstrations, reframing prompt compression as guided subsequence search. The authors report that the resulting syntactically and semantically degraded 'gibberish' prompts can improve task performance across classification, style transfer, jailbreaking, multi-choice QA, and math reasoning, often matching or exceeding learned prompt optimizers such as RLPrompt, EvoPrompt, and Promptbreeder, with competitive runtime. The paper also analyzes label-word retention in pruned prompts and includes a limitations study on template sensitivity.

Significance. If the central phenomenon is established, the paper makes a useful contribution: it provides controlled evidence that simple token-level pruning of ICL demonstrations can improve test performance across diverse models and tasks, and it connects prompt compression with prompt optimization through an evolutionary-search lens. The evaluation is broad (six classification datasets, several model families, generation and reasoning tasks), the validation/test separation is clearly stated, and the appendix includes detailed hyperparameters and a candid discussion of template sensitivity. The main weakness is that the paper's headline universal claims ('always matches or surpasses', 'consistently achieves') are contradicted by its own tables; the contribution is better characterized as 'pruning can yield strong results in many settings' rather than as a universal superiority result.

major comments (3)
  1. [Abstract; Section 4.1; Table 10] The abstract states that the gibberish 'always matches or surpasses state-of-the-art automatic prompt optimization techniques', and Section 4.1 concludes that PROMPTQUINE 'is able to match or surpass state-of-the-art performance across settings'. Table 10 directly contradicts this: in the GPT-2 block, PROMPTQUINE averages 58.6% versus RLPrompt's 58.7%, and it loses on SST-2 (77.2 vs 79.2) and AG's News (66.7 vs 75.3). Table 17 also shows PROMPTQUINE underperforming TAPruning on GSM8K for Mistral-It (45.7 vs 45.8) and Llama3-8B-It (76.4 vs 77.1). Since the 'always' and 'consistently' qualifiers are falsified by the paper's own results, these claims should be replaced with setting-dependent claims, and the paper should discuss the conditions under which pruning fails.
  2. [Section 5.1; Table 4] The template-sensitivity study in Section 5.1 shows that the pruning benefit is highly unstable: for SNLI, test accuracy of the pruned prompt fluctuates by 14.3 points (60.8 to 75.1) across three templates, with a standard deviation of 7.2. This is difficult to reconcile with the earlier claim in Section 4 that PROMPTQUINE 'consistently achieves improved results over TAPruning'. The limitation should be integrated into the abstract and conclusion, and the headline claims should be conditioned on template and seed.
  3. [Section 4.1; Appendix D.3] The two-stage selection procedure (ranking by the piecewise fitness in Equation 3 on 8 or 32 samples, then re-ranking the top k% by 200-sample validation accuracy) selects a final prompt from 10,000 searched candidates using a small validation set. Because the same validation set guides both stages, the reported test improvements partly reflect selection on validation rather than only the intrinsic pruning effect. The paper acknowledges this concern in Appendix D.3 ('validation accuracy can overfit to a narrow slice of examples'), but it does not quantify the selection bias; a validation-size or validation-split ablation would strengthen the 'low-data regime' claim.
minor comments (5)
  1. [Section 4.1] The reference to 'Appendix 4.1' does not match any appendix; this should be Appendix D.3.
  2. [Table 11] Table 11 cites Promptbreeder as '(Guo et al., 2023)'; the correct source is Fernando et al. (2024).
  3. [Introduction] In Section 1, 'could becosme more pronounced' is a typo for 'could become more pronounced'.
  4. [Abstract; Section 3.4] The 'self-replicating' and 'quine' framing is used throughout, but the implemented algorithm is a standard genetic algorithm with copy-then-mutate; the paper should either define the term operationally or explicitly note that it is used metaphorically, to avoid overclaiming novelty.
  5. [Table 1] Many of the reported improvements are within one standard deviation of the baseline (e.g., SST-2: 96.2±0.2 vs. 95.9±0.6 for ICL); a statistical significance analysis, or at least a statement about the stability of the improvements across seeds, would make the comparisons easier to interpret.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the evolved prompts are selected on held-out validation and scored on a separate test split, so the reported gains are not fitted constants; the only author-overlapping citation (RLPrompt's reward function) is a borrowed fitness proxy, not a load-bearing derivation.

full rationale

The paper's central derivation chain is an empirical search-and-evaluate loop: PROMPTQUINE mutates binary token masks, evaluates fitness via Equation 3 (a piecewise label-probability reward from Deng et al. 2022) on held-out validation samples, re-ranks elite prompts by validation accuracy, and then reports task performance on an official test split. Because selection and final evaluation use disjoint data, the claimed improvements are not equal to the fitness function by construction. Equation 3 is an input proxy, not the output claim, and the paper explicitly acknowledges its imperfection in Appendix D.3 ('it can still be exploited due to its imperfect design') and compensates with a validation-based re-ranking stage; this is a standard fit/select-then-evaluate protocol, not circularity. The one author-overlapping citation is RLPrompt (Deng et al., 2022), which contributes the reward formula and templates; borrowing a published fitness function does not make the pruning result self-defined, and the comparison numbers are re-run rather than imported. The paper's own limitations—Section 5.1's SNLI test fluctuation of 14.3%, Appendix D.6's cases where PROMPTQUINE underperforms TAPruning, and Table 10's GPT-2 average of 58.6 versus RLPrompt's 58.7—are correctness/robustness concerns about the strength of the 'always matches or surpasses' claim, not circularity, because they are empirical outcomes rather than assumptions embedded in the derivations. No step in the derivation reduces to its own input; therefore the circularity score is 0.

Assumptions & free parameters 10 free parameters · 3 assumptions · 0 invented entities

No new physical or ontological entities are introduced. The results rest on a stack of hand-tuned hyperparameters, proxy fitness functions, and the assumption that small validation sets predict test performance; the paper is transparent about several of these, particularly the imperfect fitness design in Appendix D.3.

free parameters (10)
  • TAPruning threshold δ = 96 percent of current best performance (example).
    Gates whether a suboptimal candidate is accepted during hill climbing (Algorithm 1); chosen by hand and affects which prompts survive.
  • Mutation rate set = One, two, three, or four bits flipped.
    Balances exploration and exploitation in PROMPTQUINE; selected by trial and error and listed in Appendix D.1 Table 9.
  • Population size = 30 individuals.
    GA hyperparameter from Table 9; larger values are suggested if compute allows.
  • Offspring size = 50 offspring per generation.
    GA hyperparameter from Table 9; controls how many new prompts compete each generation.
  • Tournament selection ratio = 0.2, the fraction of population sampled for selection.
    Affects selection pressure in the genetic algorithm (Appendix D.1).
  • Number of iterations = 10,000 for 1-shot SSGA; 100,000 for 4-shot GGA.
    Search budget; results are reported at convergence or budget limit, but the budget is chosen by the authors.
  • Minimal prompt length threshold = 15 tokens.
    Termination condition for the evolutionary loop (Algorithm 3, Table 9).
  • Elite selection threshold k = 10 for binary tasks, 5 for others.
    Controls how many top-fitness prompts are re-ranked on validation accuracy; calibrated by trial and error (Appendix D.3).
  • Fitness estimation sample size = 8 samples for classification, 100 for style transfer.
    Number of examples used to score each candidate prompt during search; chosen from preliminary experiments (Appendix D.3, D.4).
  • Reward scaling constants λ1, λ2 = 180 and 200.
    Weights in the piecewise fitness function (Eq. 3); adopted from Deng et al. (2022), not fitted here, but classification search depends on them.
assumptions (3)
  • domain assumption Fixed-order subsequence search space
    The paper optimizes only deletions while preserving token order (Section 2.3), so the claimed gains are for pruning, not for reordering or insertion.
  • domain assumption Validation proxy generalizes to test
    Search and re-ranking use a 200-sample held-out set, or 8 or 100 samples for fitness; the paper assumes this proxy selects prompts that also do well on the official test set, and acknowledges in Appendix D.3 that the fitness function is imperfect.
  • domain assumption Jailbreak metrics are valid measures of attack success
    ASR is measured by exact-match strings and Llama-Guard-3 judgments; if these metrics misclassify harmful outputs, the reported attack success rates would not reflect real-world jailbreak efficacy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evolving Prompts In-Context: An Open-ended, Self-replicating Perspective." pith.science (2026). https://pith.science/paper/4UX77TDU

@misc{pith2026250617930,
  author       = {Pith},
  title        = {Pith review of: Evolving Prompts In-Context: An Open-ended, Self-replicating Perspective},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4UX77TDU}},
  note         = {Machine review of arXiv:2506.17930}
}
read the original abstract

We propose a novel prompt design paradigm that challenges conventional wisdom in large language model (LLM) prompting. While conventional wisdom prioritizes well-crafted instructions and demonstrations for in-context learning (ICL), we show that pruning random demonstrations into seemingly incoherent "gibberish" can remarkably improve performance across diverse tasks. Notably, the "gibberish" always matches or surpasses state-of-the-art automatic prompt optimization techniques, achieving substantial gains regardless of LLM alignment. Nevertheless, discovering an effective pruning strategy is non-trivial, as existing attribution methods and prompt compression algorithms fail to deliver robust results, let alone human intuition. In terms of this, we propose a self-discover prompt optimization framework, PromptQuine, an evolutionary search framework that automatically searches for the pruning strategy by itself using only low-data regimes. Much like the emergent complexity in nature--such as symbiosis and self-organization--arising in response to resource constraints, our framework evolves and refines unconventional yet highly effective prompts by leveraging only the tokens present within the context. We demonstrate its effectiveness across classification, multi-choice question answering, generation and math reasoning tasks across LLMs, while achieving decent runtime efficiency. We hope our findings can guide mechanistic studies on in-context learning, and provide a call to action, to pave the way for more open-ended search algorithms for more effective LLM prompting.

Figures

Figures reproduced from arXiv: 2506.17930 by the authors.

Figure 1
Figure 1. Optimization challenges in our ICL-initialized landscape using Llama-3-8B-Instruct for subjectivity classification. Left: Randomizing pruning order in hill-climbing search leads to vary￾ing task performance, highlighting the multimodal nature. Middle: Evolutionary search (ES) outperforms random search (RS) in iden￾tifying high-quality solutions, with TAPruning result as a dashed line. Right: Relative success rate of… view at source ↗
Figure 4
Figure 4. (right). This operation further largely degrades per￾formance, highlighting the importance of preserving input￾label format as in standard ICL (Min et al., 2022). Although most of our findings so far are consistent with the findings on conventional ICL, there are still some spe￾cial cases where prompt instances violate the aggregated findings discussed above (e.g., SNLI in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Pruning-based prompting performance using a variety of methods (Top for attribution scores guided pruning, and Bottom for attention weights guided pruning). representational similarity may indicate task performance, as recent work (Shen et al., 2023a; Deutch et al., 2024) invalidate the effectiveness of such hypothesis in generalizing to practical NLP tasks for LLMs. Please note that the algorithms below are all ins… view at source ↗
Figures from the paper (4 more)
Figure 6
Figure 6. Figure 6: Additional results on both Llama-3-8B base and instruct models, revealing the complex, multimodal nature of the ICL search landscape. further corroborating our findings in the main paper. 29 [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: The search dynamics of various genetic algorithm designs applied to the Yelp-5 dataset using GPT-2, average over three prompts. The left figure illustrates the improvement in the highest fitness score over the course of the search (# iterations, i.e., the number of pro…
Figure 8
Figure 8. Figure 8: Task performance when increasing the shots in the ICL prompts. Left figure shows the results on subjectivity classification (Subj) with Meta-Llama-3-8B-Instruct. Right figure shows the results on multi-choice question answering dataset (PIQA) with Meta￾Llama-3-8B-Instr…
Figure 9
Figure 9. Figure 9: Changes in (unpruned & pruned) prompting performance on GPT-2 when labels are removed (left & middle) or even the complete outputs are removed (right) [PITH_FULL_IMAGE:figures/full_fig_p048_9.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Probabilistic Concept-Aware Steering for Trustworthy LLM Inference

    cs.AI 2026-05 reject novelty 4.0 of 10

    PCS improves steering direction accuracy by adaptively sampling the intervention coefficient from a cosine-similarity-conditioned Gaussian, but its evaluation is partly circular because the optimal coefficient is chos...

Reference graph

Works this paper leans on

16 extracted references · 3 canonical work pages · cited by 1 Pith paper

  1. [5]

    doi: 10.18653/v1/2021.emnlp-main.243

    Association for Computational Linguistics. doi: 10.18653/v1/2021.emnlp-main.243. URL https:// aclanthology.org/2021.emnlp-main.243. Li, C., Farkhoor, H., Liu, R., and Yosinski, J. Measur- ing the intrinsic dimension of objective landscapes. In International Conference on Learning Representations, 2018a. Li, J., Chen, X., Hovy, E., and Jurafsky, D. Visuali...

    arXiv 2021

  2. [8]

    doi: 10.18653/v1/2022.emnlp-main.759

    Association for Computational Linguistics. doi: 10.18653/v1/2022.emnlp-main.759. URL https:// aclanthology.org/2022.emnlp-main.759. Ngo, R., Chan, L., and Mindermann, S. The alignment problem from a deep learning perspective.arXiv preprint arXiv:2209.00626, 2022. Nisioti, E., Glanois, C., Najarro, E., Dai, A., Meyerson, E., Pedersen, J. W., Teodorescu, L....

    arXiv 2022

  3. [15]

    Zheng, L., Chiang, W.-L., Sheng, Y ., Zhuang, S., Wu, Z., Zhuang, Y ., Lin, Z., Li, Z., Li, D., Xing, E., et al

    URL https://proceedings.mlr.press/ v139/zhao21c.html. Zheng, L., Chiang, W.-L., Sheng, Y ., Zhuang, S., Wu, Z., Zhuang, Y ., Lin, Z., Li, Z., Li, D., Xing, E., et al. Judging llm-as-a-judge with mt-bench and chatbot arena.Ad- vances in Neural Information Processing Systems, 36: 46595–46623, 2023. Zheng, X., Pang, T., Du, C., Liu, Q., Jiang, J., and Lin, M...

    2023

  4. [55]

    emnlp-main.55

    URL https://aclanthology.org/2020. emnlp-main.55. Kwon, W., Li, Z., Zhuang, S., Sheng, Y ., Zheng, L., Yu, C. H., Gonzalez, J., Zhang, H., and Stoica, I. Efficient memory management for large language model serving with pagedattention. InProceedings of the 29th Sym- posium on Operating Systems Principles, pp. 611–626, 2023. Labrou, Y . and Finin, T. Yahoo...

    work page 2020

  5. [60]

    acl-long.60

    URL https://aclanthology.org/2022. acl-long.60. Souly, A., Lu, Q., Bowen, D., Trinh, T., Hsieh, E., Pandey, S., Abbeel, P., Svegliato, J., Emmons, S., Watkins, O., et al. A strongreject for empty jailbreaks.arXiv preprint arXiv:2402.10260, 2024. Stanley, K. O. and Lehman, J.Why Greatness Cannot Be Planned: The Myth of the Objective. Springer, 2015. Stanle...

    arXiv 2022

  6. [155]

    emnlp-main.155/

    URL https://aclanthology.org/2022. emnlp-main.155/. Zhang, S., Dong, L., Li, X., Zhang, S., Sun, X., Wang, S., Li, J., Hu, R., Zhang, T., Wu, F., et al. Instruction tuning for large language models: A survey.arXiv preprint arXiv:2308.10792, 2023. Zhang, X., Zhao, J., and LeCun, Y . Character-level convolu- tional networks for text classification.Advances ...

    arXiv 2022

  7. [191]

    naacl-main.191

    URL https://aclanthology.org/2022. naacl-main.191. Russell, S. J. and Norvig, P.Artificial intelligence: a modern approach. Pearson, 2016. Sareni, B. and Krahenbuhl, L. Fitness sharing and niching methods revisited.IEEE transactions on Evolutionary Computation, 2(3):97–106, 1998. Schick, T. and Sch ¨utze, H. Exploiting cloze-questions for few-shot text cl...

    arXiv 2022

  8. [247]

    findings-acl.247

    URL https://aclanthology.org/2023. findings-acl.247. Daras, G. and Dimakis, A. Discovering the hidden vocab- ulary of dalle-2. InNeurIPS 2022 Workshop on Score- Based Methods, 2022. Dasgupta, I., Grant, E., and Griffiths, T. Distinguishing rule and exemplar-based generalization in learning systems. InInternational Conference on Machine Learning, pp. 4816–...

    work page 2023

Show all 16 references
  1. [346]

    emnlp-main.346

    URL https://aclanthology.org/2020. emnlp-main.346. Shrikumar, A., Greenside, P., Shcherbina, A., and Kundaje, A. Not just a black box: Learning important features through propagating activation differences.arXiv preprint arXiv:1605.01713, 2016. Simonyan, K. Deep inside convolu...

    2020 arXiv

  2. [353]

    acl-long.353

    URL https://aclanthology.org/2021. acl-long.353. Li, Y ., Dong, B., Guerin, F., and Lin, C. Compressing context to enhance inference efficiency of large language models. In Bouamor, H., Pino, J., and Bali, K. (eds.), Proceedings of the 2023 Conference on Empirical Meth- ods in...

    2021 doi

  3. [391]

    emnlp-main.391

    URL https://aclanthology.org/2023. emnlp-main.391. Liao, C., Zheng, Y ., and Yang, Z. Zero-label prompt selec- tion.arXiv preprint arXiv:2211.04668, 2022. Lightman, H., Kosaraju, V ., Burda, Y ., Edwards, H., Baker, B., Lee, T., Leike, J., Schulman, J., Sutskever, I., and Cobb...

    2023 arXiv

  4. [793]

    secret language

    URL https://aclanthology.org/2022. emnlp-main.793. Zou, A., Wang, Z., Kolter, J. Z., and Fredrikson, M. Uni- versal and transferable adversarial attacks on aligned lan- guage models.arXiv preprint arXiv:2307.15043, 2023. 18 Evolving Prompts In-Context: An Open-ended, Self-repl...

    2020 arXiv

  5. [2004]

    URL https: //aclanthology.org/P04-1035

    doi: 10.3115/1218955.1218990. URL https: //aclanthology.org/P04-1035. Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. Bleu: a method for automatic evaluation of machine translation. In Isabelle, P., Charniak, E., and Lin, D. (eds.),Proceed- ings of the 40th Annual Meeting o...

  6. [2013]

    URL https://aclanthology.org/D13-1170

    Association for Computational Linguistics. URL https://aclanthology.org/D13-1170. Sorensen, T., Robinson, J., Rytting, C., Shaw, A., Rogers, K., Delorey, A., Khalil, M., Fulda, N., and Wingate, D. An information-theoretic approach to prompt engi- neering without ground truth l...

    2022

  7. [2021]

    doi: 10.18653/v1/2021.emnlp-main.599

    Association for Computational Linguistics. doi: 10.18653/v1/2021.emnlp-main.599. URL https:// aclanthology.org/2021.emnlp-main.599. Deng, M., Wang, J., Hsieh, C.-P., Wang, Y ., Guo, H., Shu, T., Song, M., Xing, E., and Hu, Z. RLPrompt: Optimizing discrete text prompts with rei...

    2021 arXiv

  8. [2022]

    doi: 10.18653/v1/2022.naacl-main.266

    Association for Computational Linguistics. doi: 10.18653/v1/2022.naacl-main.266. URL https:// aclanthology.org/2022.naacl-main.266. Kojima, T., Gu, S. S., Reid, M., Matsuo, Y ., and Iwasawa, Y . Large language models are zero-shot reasoners.Ad- vances in neural information pro...

    2022 doi

Pith tools

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