Pith. sign in

REVIEW 4 major objections 5 minor 66 references

PromptAL: Sample-Aware Dynamic Soft Prompts for Few-Shot Active Learning

T0 review · 4 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 PromptAL claims that per-sample dynamic soft prompts built from unlabeled data can correct the decision boundary and make few-shot active learning more label-efficient than nine baselines.

desk verdict The idea is a real combination, but the paper hides the training of its core prompt modules behind a forward-pass algorithm and overclaims on its own Table 4. read the letter →

arxiv 2507.16424 v1 pith:PMFLAF6Y submitted 2025-07-22 cs.CL cs.LG

Hui Xiang , Jinqiao Shi , Ting Zhang , Xiaojie Zhao , Yong Liu , Yong Ma This is my paper
classification cs.CLcs.LG
keywords activelearningfew-shotsoftpromptssample-awaredecisionboundaryalignmentuncertaintysamplingdiversitytextclassification
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

PromptAL is a framework for few-shot active learning in text classification. Its central claim is that the decision boundary used to choose which unlabeled examples to label is unreliable when only a few labeled examples exist, and that this can be fixed by using the unlabeled pool itself to build sample-aware dynamic soft prompts for a masked-language model. With the adjusted predictions, the method ranks unlabeled examples by a weighted combination of entropy-based uncertainty, global diversity from clustering in a knowledge-feature space, and local diversity measured against the labeled set. Across six in-domain datasets and three out-of-domain sentiment datasets, the paper reports higher average accuracy than nine baselines, with the largest gains on datasets with many classes. If the claim holds, few-shot annotation budgets can be spent on examples that better represent the target distribution.

What carries the argument

The central object is the sample-aware dynamic soft prompt $P(x) = \mathrm{MultiHead}(\mathrm{concat}[T, S(x)])$, where $T$ is a learnable task prompt shared by all inputs and $S(x)$ is a sample prompt produced by the generator $f$ (an encoder plus a two-layer MLP). Feeding $P(x)$ into a RoBERTa-style masked-language model shifts the predicted probability of the verbalizer label word, and the paper treats that shift as a correction of the decision boundary toward the target distribution. The same mask-token representation $h_{\langle\mathrm{MASK}\rangle}$ defines the knowledge feature space used for global-diversity clustering, while a KNN distance to the labeled set gives local diversity; these feed the joint score $S(x) = \lambda U(x) + (1-\lambda) D(x)$ that ranks candidates.

What would settle it

Rerun the ten-round active-learning protocol with the sample-prompt generator and attention weights frozen at random initialization while keeping all other settings identical; if accuracy stays at the published level, the sample-aware mechanism is not doing the work, and if the Section 5.8 JS-divergence gap disappears, the decision-boundary story is unsupported.

Watch

Extended reading notes

Core claim

The paper's central claim is that unlabeled samples should be used not merely to score informativeness but to actively reshape the model's predictive distribution. PromptAL does this by generating a per-sample dynamic soft prompt from a shared task prompt and a sample-specific prompt, feeding it to a masked-language model before the mask token, and using the resulting calibrated distribution for entropy-based uncertainty. Diversity is then enforced both globally, by k-means++ clustering on the mask-token knowledge feature, and locally, by penalizing closeness to labeled training examples. The paper argues that this produces a query set whose category balance, representativeness, uncertainty, and diversity are better than those of uncertainty-only, diversity-only, and earlier hybrid baselines, and that the selected samples accelerate convergence toward the target distribution.

Load-bearing premise

The load-bearing premise is that the learnable components that build the sample-aware prompt (the task prompt, the sample-prompt generator, and the attention fusion) are actually trained in a way that improves the decision boundary; the paper does not describe that training procedure, so if those parameters are not optimized, the reported gains cannot be attributed to sample-awareness.

Editorial extensions

If this is right

  • On high-class-count datasets such as TREC, DBPedia, and Yahoo, PromptAL reaches comparable accuracy in roughly three active-learning iterations where entropy sampling needs ten, so the same annotation budget buys faster convergence.
  • The reported gains over the hard-prompt baseline Patron (about two percentage points on DBPedia) imply that per-sample prompt adaptation, not just prompt tuning, is responsible for part of the improvement.
  • The calibration module is load-bearing: removing it degrades performance, so any successful extension must keep a correction for the model's word-prediction bias.
  • On three out-of-domain sentiment benchmarks, PromptAL reports the highest accuracy on two datasets and beats Patron by 3.16 percent on average, suggesting that the selected samples transfer better.
  • The JS-divergence experiment reports that adding sample-aware information moves the model's distribution closer to a fully trained model's distribution on all six datasets, directly testing the paper's motivation.

Reading between the lines

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

  • The unlabeled-pool information is used in a transductive way, reshaping the model before any new labels arrive; if the effect is real, other few-shot active-learning methods could adopt the same distribution-correction-before-querying step without replacing their scoring functions.
  • The paper leaves the training objective for the task prompt, the sample-prompt generator, and the attention fusion unspecified; a reader cannot tell from the text whether these are optimized during fine-tuning or on the unlabeled pool, so replication currently depends on the released code.
  • The mask-token knowledge feature is a natural interface between prompt tuning and active learning; it could be tested on other masked-language backbones or in multilingual few-shot settings to see whether the gains transfer.
  • A cleaner causal test would compare PromptAL against a version whose dynamic prompt is replaced by a random projection of the same sample features, isolating whether the decision-boundary mechanism or the extra features drive the gains.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes PromptAL, an active-learning framework for few-shot text classification. The method generates sample-aware dynamic soft prompts by fusing a learnable task prompt and a per-sample prompt (produced by an encoder plus MLP) through multi-head self-attention, uses the resulting masked-language-model distribution for calibrated entropy-based uncertainty, applies k-means++ clustering on mask-token knowledge features for global diversity, and uses KNN distance to labeled data for local diversity, then selects one sample per cluster by a weighted joint score. Experiments on six in-domain and three out-of-domain datasets compare against nine baselines, with ablations, hyperparameter analysis, a distribution-alignment study, and an LLM comparison. The paper claims that PromptAL consistently outperforms all baselines across all six tasks and demonstrates superior generalization.

Significance. If the central mechanism works as described, PromptAL would contribute a novel path for active learning in few-shot settings: using unlabeled samples to reshape the model's predictive distribution before querying, rather than relying only on labeled data. The paper is commendable for attempting statistical significance tests, OOD evaluation, ablations of each component, a distribution-alignment experiment, computational-cost analysis, and release of code. These are strengths. However, the core prompt-generation components are not specified enough to be reproduced, the headline 'consistently outperforms' claim is contradicted by the paper's own Table 4, and the hyperparameter selection appears to use the test sets, all of which are load-bearing for the reported gains. The significance of the contribution is therefore conditional on resolving these issues.

major comments (4)
  1. [§4.1, Algorithm 2, §5.1.3, §5.6.1] The training procedure for the core dynamic-soft-prompt mechanism is missing. Section 4.1 defines P(x) = MultiHead(concat[T, S(x)]) with T a learnable task prompt and S(x) generated by encoder E and a two-layer MLP, but neither the main text, Algorithm 2, nor Section 5.1.3 states the objective, loss, optimizer configuration, or which parameters are updated when Algorithm 1 performs 'Mi ← tune M0 on Dtrain'. Algorithm 2 is purely a forward pass over the unlabeled pool. The ablation in Section 5.6.1 says that removing the sample-specific prompt makes 'the prompt revert to a fixed soft prompt after training', which implies that T, f, and the attention fusion are trained, but no training recipe is given. As written, the paper is not reproducible; if these components are untrained, P(x) is an untrained transformation and cannot support the claimed distribution-alignment effect. This is the load-bearing gap behind the paper's main claim.
  2. [§5.3, Table 4, Table 5] The statement in Section 5.3 that 'PromptAL consistently outperforms all other methods across all six tasks' is contradicted by Table 4, where Patron is higher than PromptAL on AGNews (87.13 vs. 86.84), and by Table 5, where the paired t-tests are not significant on AGNews for BADGE (p=0.159), CAL (p=0.270), and Patron (p=0.544), and on IMDB for BERT-KM (p=0.101). Section 5.4 later acknowledges that PromptAL does not reach statistical significance against four baselines on AGNews and IMDB, so the abstract's 'superior performance over nine baselines' and Section 5.3's 'consistently outperforms' phrasing overstate the evidence. The claims should be qualified to reflect the datasets and baselines where the advantage is actually significant.
  3. [§5.7, Table 4, Figures 7–9] The hyperparameter analysis appears to select configuration values on the same test sets used for the final reported results. Section 5.7 reports grid searches for the attention-head count, task-prompt size m, sample-prompt size n, and joint weight λ, with accuracy curves for TREC and IMDB that appear to be test accuracies, and Section 5.1.3 fixes the final values in Table 3 before Table 4 is presented. No held-out validation set or nested validation procedure is described. If the hyperparameters were tuned on the test sets, the reported margins in Table 4 are optimistically biased. The authors should either describe a proper validation protocol or re-run the comparisons with hyperparameters selected on development data.
  4. [§4.2, Eq. (5)] The construction of the support set R used for contextualized-prior calibration is underspecified. The text says 'We first construct a support set R by selecting the top k samples with the highest p(yi|x) for each label yi ∈ Y' but does not state whether these samples come from the labeled training set, the unlabeled pool, or some other source, nor what labels are used for samples without ground-truth labels. Since the calibration step in Eq. (6) directly affects the uncertainty scores that drive query selection, the origin of R and the labeling procedure must be specified for the method to be reproducible and for the calibration effect to be interpretable.
minor comments (5)
  1. [Table 3, §4.1, Fig. 8] The notation for prompt sizes is inconsistent: Section 4.1 defines T ∈ R^{m×d} and S(x) ∈ R^{n×d}, while Table 3 swaps the names ('vector number of task-specific prompt n 4' and 'vector number of sample-specific prompt m 1'), and Figure 8 labels the task-specific size as m and sample-specific size as n, creating confusion about which size is 4 and which is 1.
  2. [Algorithm 2] Algorithm 2 uses the symbol Q both for the cluster set and the final query set, which is confusing; renaming the clusters (e.g., C) would make the two selection steps clearer. The line 'Q ←k-MEANS++ cluster batch set of D' also contains a formatting/notation error.
  3. [Fig. 4 and throughout §5.6.1] The legend labels in Figure 4 contain a typo: 'w/o T ask-Specific Prompt' should read 'w/o Task-Specific Prompt'. The same legend spacing issue appears in the figure caption text.
  4. [§8, Conclusions] The conclusions state that PromptAL 'outperforms seven baseline methods', while Section 5.2 and the abstract refer to nine baselines; the count should be made consistent.
  5. [§5.5, Table 6] The OOD comparison says PromptAL 'outperforms the latest method, Patron, by 3.16% across all three OOD datasets', but Table 6 shows PromptAL does not beat Patron on IMDB-Counter (92.8 vs. 89.8? Actually Patron is 89.8, so this is fine; the sentence is accurate). Clarify that the 3.16% is an average margin, since per-dataset margins differ.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PromptAL's reported gains rest on empirical benchmark comparisons rather than on a derivation that reduces to its own inputs.

full rationale

PromptAL's central claim is an empirical comparison: Section 5.3 reports accuracy curves and Table 4 reports average accuracy against nine baselines, so the result is a benchmark outcome rather than a derived theorem. The query score S(x) = lambda * U(x) + (1 - lambda) * D(x) (Eq. 9) is assembled from entropy (Eq. 7), KNN distance (Eq. 8), and k-means++ cluster assignment, and none of these quantities is defined in terms of the final test accuracy or of the labels of the queried samples; therefore the reported improvement is not forced by construction. The dynamic soft prompt P(x) = MultiHead(concat[T, S(x)]) (Eq. 1) is an input transformation whose parameters are learnable, and although the paper omits a description of how T, f, and the attention weights are optimized, an omitted training procedure is a reproducibility defect rather than a circular equation. The hyperparameter choices in Section 5.7 (m = 4, n = 1, lambda = 0.9, four attention heads) are selected configurations and are not presented as predictions of test accuracy; if the grid search was conducted on test accuracy, that is a selection-bias and correctness concern, not circularity. The distribution-alignment experiment in Section 5.8 compares the JS divergence of two model distributions against a fully trained model's test distribution; this supports the motivation but is not a derivation whose conclusion is identical to its premise. I found no self-citation chain, no imported uniqueness theorem, and no renaming of a known result presented as a new organization. The appropriate finding is therefore no significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on the prompt mechanism, the chosen feature space, and multiple hand-set hyperparameters. The training procedure for the prompt modules is unstated, and several configuration values are tuned on the same test sets used for comparison.

free parameters (7)
  • Joint score weight lambda = 0.9
    Grid search over [0.1, 0.5, 0.9, 1.0] using test accuracy on TREC and IMDB (Section 5.7, Figure 9); no separate validation split is described.
  • Task-specific prompt size m = 4
    Grid search over {2, 4, 8, 16} using test accuracy on TREC and IMDB (Section 5.7, Figure 8).
  • Sample-specific prompt size n = 1
    Grid search over {1, 4, 8, 16} using test accuracy on TREC and IMDB (Section 5.7, Figure 8).
  • Number of self-attention heads = 4
    Compared 1, 4, and 8 heads on TREC (Section 5.7, Figure 7); 4 heads selected.
  • Support set size k = 100
    Used to estimate the contextualized prior in Eq. 5; no sensitivity analysis is provided.
  • Local diversity neighbor count k' = 10
    Number of nearest labeled samples used in Eq. 8; no sensitivity analysis is provided.
  • Prompt generator hidden dimension = 256
    Set for the MLP in the sample prompt generator f (Section 5.1.3); no ablation is reported.
assumptions (5)
  • standard math Multi-head self-attention, K-means++, KNN, and entropy are used as prescribed by their standard definitions.
    Equations 1-2 and Sections 4.3.1-4.3.2 use standard tools; no modification requiring proof is introduced.
  • domain assumption The MASK-token hidden state h_MASK from the dynamically prompted PLM is a 'knowledge feature' that captures task-relevant information better than [CLS] or SimCSE embeddings.
    Section 4.3.1 and Section 5.6.3; the ablation supports this empirically, but no theoretical justification is given, and the feature is generated by the same prompt mechanism being evaluated.
  • domain assumption Unlabeled data can be used to construct soft prompts that shift the predictive distribution toward the target distribution without requiring labels.
    This is the core motivation in Section 1 and is tested only indirectly in Section 5.8 using JS divergence against a fully trained target model.
  • ad hoc to paper A support set R for contextualized prior calibration is available from a specified source, but the paper does not state whether R comes from labeled or unlabeled data.
    Section 4.2, Eq. 5 says 'top k samples' but does not specify the pool; the choice affects calibration and all downstream uncertainty scores.
  • ad hoc to paper The number of global clusters is set equal to the acquisition batch size b, and exactly one sample is selected from each cluster.
    Algorithm 2, lines 11-12; no justification is provided for why b clusters with one sample per cluster is the right decomposition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PromptAL: Sample-Aware Dynamic Soft Prompts for Few-Shot Active Learning." pith.science (2026). https://pith.science/paper/PMFLAF6Y

@misc{pith2026250716424,
  author       = {Pith},
  title        = {Pith review of: PromptAL: Sample-Aware Dynamic Soft Prompts for Few-Shot Active Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PMFLAF6Y}},
  note         = {Machine review of arXiv:2507.16424}
}
read the original abstract

Active learning (AL) aims to optimize model training and reduce annotation costs by selecting the most informative samples for labeling. Typically, AL methods rely on the empirical distribution of labeled data to define the decision boundary and perform uncertainty or diversity estimation, subsequently identifying potential high-quality samples. In few-shot scenarios, the empirical distribution often diverges significantly from the target distribution, causing the decision boundary to shift away from its optimal position. However, existing methods overlook the role of unlabeled samples in enhancing the empirical distribution to better align with the target distribution, resulting in a suboptimal decision boundary and the selection of samples that inadequately represent the target distribution. To address this, we propose a hybrid AL framework, termed \textbf{PromptAL} (Sample-Aware Dynamic Soft \textbf{Prompts} for Few-Shot \textbf{A}ctive \textbf{L}earning). This framework accounts for the contribution of each unlabeled data point in aligning the current empirical distribution with the target distribution, thereby optimizing the decision boundary. Specifically, PromptAL first leverages unlabeled data to construct sample-aware dynamic soft prompts that adjust the model's predictive distribution and decision boundary. Subsequently, based on the adjusted decision boundary, it integrates uncertainty estimation with both global and local diversity to select high-quality samples that more accurately represent the target distribution. Experimental results on six in-domain and three out-of-domain datasets show that PromptAL achieves superior performance over nine baselines. Our codebase is openly accessible.

Figures

Figures reproduced from arXiv: 2507.16424 by the authors.

Figure 1
Figure 1. Motivation behind PromptAL. cal distribution with the target distribution, thereby optimizing the decision boundary. Specifically, PromptAL first leverages unlabeled data to construct sample-aware dynamic soft prompts that adjust the model’s predictive distri￾bution and decision boundary. Subsequently, based on the adjusted decision boundary, it integrates uncertainty estimation with both global and local diversity … view at source ↗
Figure 2
Figure 2. Principle of PromptAL Framework. The process consists of four main steps: [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. In-domain test accuracy during AL iterations for different query strategies. [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Results of the PromptAL module ablation analysis. [PITH_FULL_IMAGE:figures/full_fig_p021_4.png]
Figure 5
Figure 5. Figure 5: Results of different fusion mechanisms. We assessed different fusion mechanisms for integrating task and sample prompts. In PromptAL, the self-attention mechanism is substituted with two alternatives: the Hadamard Product and Element-wise Addition. In both approaches, …
Figure 6
Figure 6. Figure 6: Results of different feature spaces for global diversity. [PITH_FULL_IMAGE:figures/full_fig_p023_6.png]
Figure 7
Figure 7. Figure 7: Impact of attention head num￾bers. 0 2 4 6 8 10 12 Task Prompt Params (K) Parameters 2 4 8 16 Task-Specific Prompt Size m 75 80 85 90 95 Accuracy (in %) TREC IMDB 0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 Prompt Generator Params(M) Parameters 1 4 8 16 Sample-Specific Prompt Size…
Figure 10
Figure 10. Figure 10: Degree of model fit to the target distribution. The model with unlabeled [PITH_FULL_IMAGE:figures/full_fig_p024_10.png]
Figure 11
Figure 11. Figure 11: Samples selected by Entropy and PromptAL after the first AL round on the [PITH_FULL_IMAGE:figures/full_fig_p029_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 40 canonical work pages

  1. [1]

    Zhang, E

    Z. Zhang, E. Strubell, E. Hovy, A Survey of Active Learning for Natural Language Processing, in: Y. Goldberg, Z. Kozareva, Y. Zhang (Eds.), Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, Abu Dhabi, United Arab Emirates, 2022, pp. 6166–6190.doi:10.18653/v1/ 2022.emnlp-main.414

    doi:10.18653/v1/ 2022

  2. [2]

    Culotta, A

    A. Culotta, A. McCallum, Reducing labeling effort for structured pre- diction tasks, in: AAAI, Vol. 5, 2005, pp. 746–751

    work page 2005

  3. [3]

    J. T. Ash, C. Zhang, A. Krishnamurthy, J. Langford, A. Agarwal, Deep Batch Active Learning by Diverse, Uncertain Gradient Lower Bounds, in: International Conference on Learning Representations, 2019

    work page 2019

  4. [4]

    Margatina, G

    K. Margatina, G. Vernikos, L. Barrault, N. Aletras, Active Learning by Acquiring Contrastive Examples, in: M.-F. Moens, X. Huang, L. Specia, S. W.-t. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Association for Computa- tional Linguistics, Online and Punta Cana, Dominican Republic, 2021, pp. 650–663. ...

    doi:10.18653/v1/2021.emnlp-main.51 2021

  5. [5]

    C. E. Shannon, A mathematical theory of communication, The Bell system technical journal 27 (3) (1948) 379–423

    1948

  6. [6]

    Y. Song, T. Wang, P. Cai, S. K. Mondal, J. P. Sahoo, A comprehensive survey of few-shot learning: Evolution, applications, challenges, and opportunities, ACM Comput. Surv. 55 (13s) (Jul. 2023).doi:10.1145/ 3582688. URL https://doi.org/10.1145/3582688

    doi:10.1145/3582688 2023

  7. [7]

    L. E. Dor, A. Halfon, A. Gera, E. Shnarch, L. Dankin, L. Choshen, M. Danilevsky, R. Aharonov, Y. Katz, N. Slonim, Active learning for BERT: An empirical study, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2020, pp. 7949–7962

    work page 2020

  8. [8]

    Dasgupta, Two faces of active learning, Theoretical computer science 412 (19) (2011) 1767–1781

    S. Dasgupta, Two faces of active learning, Theoretical computer science 412 (19) (2011) 1767–1781. 33

    work page 2011

Show all 66 references
  1. [9]

    M. Tang, X. Luo, S. Roukos, Active learning for statistical natural lan- guage parsing, in: Proceedings of the 40th Annual Meeting of the Asso- ciation for Computational Linguistics, 2002, pp. 120–127

    2002

  2. [10]

    Maekawa, D

    S. Maekawa, D. Zhang, H. Kim, S. Rahman, E. Hruschka, Low-resource interactive active labeling for fine-tuning language models, in: Findings of the Association for Computational Linguistics: EMNLP 2022, 2022, pp. 3230–3242

    2022

  3. [11]

    Chang, X

    E. Chang, X. Shen, H.-S. Yeh, V. Demberg, On training instance se- lection for few-shot neural text generation, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Proceedings of the 59th Annual Meeting of the Asso- ciation for Computational Linguistics and the 11th International J...

    2021

  4. [13]

    doi:10.18653/v1/2021.acl-short.2

    2021 doi

  5. [14]

    T. Shi, A. Benton, I. Malioutov, O. İrsoy, Diversity-aware batch active learning for dependency parsing, in: Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, 2021, pp. 2616–2626

    2021

  6. [15]

    Zhdanov, Diverse mini-batch active learning, arXiv preprint arXiv:1901.05954 (2019)

    F. Zhdanov, Diverse mini-batch active learning, arXiv preprint arXiv:1901.05954 (2019)

    2019 arXiv

  7. [16]

    Yuan, H.-T

    M. Yuan, H.-T. Lin, J. Boyd-Graber, Cold-start Active Learning through Self-supervised Language Modeling, in: B. Webber, T. Cohn, Y. He, Y. Liu (Eds.), Proceedings of the 2020 Conference on Em- pirical Methods in Natural Language Processing (EMNLP), Associa- tion for Computati...

    2020 doi

  8. [17]

    Köksal, T

    A. Köksal, T. Schick, H. Schuetze, MEAL: Stable and Active Learning for Few-Shot Prompting, in: H. Bouamor, J. Pino, K. Bali (Eds.), Find- ings of the Association for Computational Linguistics: EMNLP 2023, 34 Association for Computational Linguistics, Singapore, 2023, pp. 506–

    2023

  9. [18]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, K. Toutanova, Bert: Pre-training of deep bidirectional transformers for language understanding, arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  10. [19]

    Y. Yu, R. Zhang, R. Xu, J. Zhang, J. Shen, C. Zhang, Cold-Start Data Selection for Better Few-shot Language Model Fine-tuning: A Prompt- based Uncertainty Propagation Approach, in: A. Rogers, J. Boyd- Graber, N. Okazaki (Eds.), Proceedings of the 61st Annual Meeting of theAsso...

    2023 doi

  11. [20]

    Jawahar, B

    G. Jawahar, B. Sagot, D. Seddah, What does bert learn about the struc- ture of language?, in: ACL 2019-57th Annual Meeting of the Association for Computational Linguistics, 2019

    2019

  12. [21]

    Liu, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 364 (2019).arXiv:1907.11692

    Y. Liu, Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907.11692 364 (2019).arXiv:1907.11692

    2019 arXiv

  13. [22]

    Hewitt, C

    J. Hewitt, C. D. Manning, A structural probe for finding syntax in word representations, in: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), 2019, ...

    2019

  14. [23]

    Yenicelik, F

    D. Yenicelik, F. Schmidt, Y. Kilcher, How does bert capture semantics? a closer look at polysemous words, in: Proceedings of the Third Black- boxNLP Workshop on Analyzing and Interpreting Neural Networks for NLP, 2020, pp. 156–162

    2020

  15. [24]

    P. Liu, W. Yuan, J. Fu, Z. Jiang, H. Hayashi, G. Neubig, Pre-train, Prompt, and Predict: A Systematic Survey of Prompting Methods in Natural Language Processing, ACM Computing Surveys 55 (9) (2023) 195:1–195:35. doi:10.1145/3560815. 35

    2023 doi

  16. [25]

    Petroni, T

    F. Petroni, T. Rocktäschel, P. Lewis, A. Bakhtin, Y. Wu, A. H. Miller, S. Riedel, Language models as knowledge bases?, arXiv preprint arXiv:1909.01066 (2019)

    2019 arXiv

  17. [26]

    Y. Gu, X. Han, Z. Liu, M. Huang, PPT: Pre-trained prompt tun- ing for few-shot learning, in: S. Muresan, P. Nakov, A. Villavicencio (Eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Association for Computat...

    2022 doi

  18. [27]

    Schick, H

    T. Schick, H. Schütze, Exploiting Cloze-Questions for Few-Shot Text Classification and Natural Language Inference, in: P. Merlo, J. Tiede- mann, R. Tsarfaty (Eds.), Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Mai...

    2021 doi

  19. [28]

    Z. Wu, S. Wang, J. Gu, R. Hou, Y. Dong, V. Vydiswaran, H. Ma, IDPG: An Instance-Dependent Prompt Generation Method, in: M. Carpuat, M.-C. de Marneffe, I. V. Meza Ruiz (Eds.), Proceedings of the 2022 Conference of the North American Chapter of the Association for Com- putationa...

    2022 doi

  20. [29]

    X. Liu, Y. Zheng, Z. Du, M. Ding, Y. Qian, Z. Yang, J. Tang, GPT understands, too, AI Open 5 (2024) 208–215.doi:10.1016/j.aiopen. 2023.08.012

    2024 doi

  21. [30]

    Z. Zhao, E. Wallace, S. Feng, D. Klein, S. Singh, Calibrate before use: Improving few-shot performance of language models, in: International conference on machine learning, PMLR, 2021, pp. 12697–12706

    2021

  22. [31]

    A. Jain, S. Chaudhuri, T. Reps, C. Jermaine, Prompt Tuning Strikes Back: Customizing Foundation Models with Low-Rank Prompt Adap- tation, in: The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  23. [32]

    D. D. Lewis, A sequential algorithm for training text classifiers: Corri- gendum and additional data, in: Acm Sigir Forum, Vol. 29, ACM New York, NY, USA, 1995, pp. 13–19

    1995

  24. [33]

    S. Hu, N. Ding, H. Wang, Z. Liu, J. Wang, J. Li, W. Wu, M. Sun, Knowledgeable Prompt-tuning: Incorporating Knowledge into Prompt Verbalizer for Text Classification, in: S. Muresan, P. Nakov, A. Villavi- cencio(Eds.), Proceedingsofthe60thAnnualMeetingoftheAssociation for Comput...

    2022 doi

  25. [34]

    Arthur, S

    D. Arthur, S. Vassilvitskii, k-means++: the advantages of careful seed- ing, in: Proceedings of the Eighteenth Annual ACM-SIAM Symposium on Discrete Algorithms, SODA ’07, Society for Industrial and Applied Mathematics, USA, 2007, p. 1027–1035

    2007

  26. [35]

    Jiang, J

    T. Jiang, J. Jiao, S. Huang, Z. Zhang, D. Wang, F. Zhuang, F. Wei, H. Huang, D. Deng, Q. Zhang, PromptBERT: Improving BERT Sen- tence Embeddings with Prompts, in: Y. Goldberg, Z. Kozareva, Y.Zhang(Eds.), Proceedingsofthe2022ConferenceonEmpiricalMeth- ods in Natural Language Pr...

    2022 doi

  27. [36]

    A. Maas, R. E. Daly, P. T. Pham, D. Huang, A. Y. Ng, C. Potts, Learning word vectors for sentiment analysis, in: Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics: Human Language Technologies, 2011, pp. 142–150

    2011

  28. [37]

    Cover, P

    T. Cover, P. Hart, Nearest neighbor pattern classification, IEEE Trans- actions on Information Theory 13 (1) (1967) 21–27.doi:10.1109/TIT. 1967.1053964

    1967

  29. [38]

    Lehmann, R

    J. Lehmann, R. Isele, M. Jakob, A. Jentzsch, D. Kontokostas, P. N. Mendes, S. Hellmann, M. Morsey, P.Van Kleef, S. Auer, et al., Dbpedia– a large-scale, multilingual knowledge base extracted from wikipedia, Se- mantic web 6 (2) (2015) 167–195

    2015

  30. [39]

    Zhang, J

    X. Zhang, J. Zhao, Y. LeCun, Character-levelconvolutionalnetworksfor text classification, Advances in neural information processing systems 28 (2015)

    2015

  31. [40]

    X. Li, D. Roth, Learning question classifiers, in: COLING 2002: The 19th International Conference on Computational Linguistics, 2002

    2002

  32. [41]

    Y. Meng, J. Shen, C. Zhang, J. Han, Weakly-supervised hierarchical text classification, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, 2019, pp. 6826–6833. 37

    2019

  33. [42]

    Gardner, Y

    M. Gardner, Y. Artzi, V. Basmov, J. Berant, B. Bogin, S. Chen, P. Dasigi, D. Dua, Y. Elazar, A. Gottumukkala, et al., Evaluating mod- els’ local decision boundaries via contrast sets, in: Findings of the As- sociation for Computational Linguistics: EMNLP 2020, 2020, pp. 1307– 1323

    2020

  34. [43]

    Socher, A

    R. Socher, A. Perelygin, J. Wu, J. Chuang, C. D. Manning, A. Y. Ng, C. Potts, Recursive deep models for semantic compositionality over a sentimenttreebank, in: Proceedingsofthe2013ConferenceonEmpirical Methods in Natural Language Processing, 2013, pp. 1631–1642

    2013

  35. [44]

    T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cis- tac, T. Rault, R. Louf, M. Funtowicz, et al., Transformers: State-of-the- art natural language processing, in: Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System D...

    2020

  36. [45]

    Kaushik, E

    D. Kaushik, E. Hovy, Z. Lipton, Learning the difference that makes a difference with counterfactually-augmented data, in: International Con- ference on Learning Representations, 2020

    2020

  37. [46]

    Houlsby, F

    N. Houlsby, F. Huszár, Z. Ghahramani, M. Lengyel, Bayesian ac- tive learning for classification and preference learning, arXiv preprint arXiv:1112.5745 (2011)

    2011 arXiv

  38. [47]

    Loshchilov, F

    I. Loshchilov, F. Hutter, Decoupled weight decay regularization, arXiv preprint arXiv:1711.05101 (2017). arXiv:1711.05101

    2017 arXiv

  39. [48]

    T. Gao, X. Yao, D. Chen, SimCSE: Simple contrastive learning of sen- tence embeddings, in: M.-F. Moens, X. Huang, L. Specia, S. W.-t. Yih (Eds.), Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguis- tics...

    2021 doi

  40. [49]

    Sener, S

    O. Sener, S. Savarese, Active learning for convolutional neural networks: A core-set approach, in: International Conference on Learning Repre- sentations, 2018

    2018

  41. [50]

    J. Zhu, H. Wang, T. Yao, B. K. Tsou, Active learning with sampling by uncertainty and density for word sense disambiguation and text classifi- cation, in: 22nd International Conference on Computational Linguistics, Coling 2008, 2008, pp. 1137–1144

    2008

  42. [51]

    Lin, Divergence measures based on the shannon entropy, IEEE Trans- actions on Information Theory 37 (1) (1991) 145–151

    J. Lin, Divergence measures based on the shannon entropy, IEEE Trans- actions on Information Theory 37 (1) (1991) 145–151. doi:10.1109/ 18.61115

    1991

  43. [52]

    van der Maaten, G

    L. van der Maaten, G. Hinton, Visualizing data using t-sne, Journal of Machine Learning Research 9 (86) (2008) 2579–2605. URL http://jmlr.org/papers/v9/vandermaaten08a.html

    2008

  44. [53]

    J. L. Bentley, Multidimensional binary search trees used for associative searching, Communications of the ACM 18 (9) (1975) 509–517

    1975

  45. [54]

    Qwen, :, A. Yang, B. Yang, B. Zhang, B. Hui, B. Zheng, B. Yu, C. Li, D. Liu, F. Huang, H. Wei, H. Lin, J. Yang, J. Tu, J. Zhang, J. Yang, J. Yang, J. Zhou, J. Lin, K. Dang, K. Lu, K. Bao, K. Yang, L. Yu, M. Li, M. Xue, P. Zhang, Q. Zhu, R. Men, R. Lin, T. Li, T. Tang, T. Xia, ...

    2025 arXiv

  46. [55]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, A. Letman, A. Mathur, A. Schelten, A. Vaughan, et al., The llama 3 herd of models, arXiv preprint arXiv:2407.21783 (2024)

    2024 arXiv

  47. [56]

    Sahoo, A

    P. Sahoo, A. K. Singh, S. Saha, V. Jain, S. Mondal, A. Chadha, A sys- tematic survey of prompt engineering in large language models: Tech- niques and applications (2025).arXiv:2402.07927. URL https://arxiv.org/abs/2402.07927

    2025 arXiv

  48. [57]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al., Deepseek-v3 technical report, arXiv preprint arXiv:2412.19437 (2024)

    2024 arXiv

  49. [58]

    W. Kwon, Z. Li, S. Zhuang, Y. Sheng, L. Zheng, C. H. Yu, J. E. Gon- zalez, H. Zhang, I. Stoica, Efficient memory management for large lan- guage model serving with pagedattention, in: Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    2023

  50. [59]

    E. J. Hu, Y. Shen, P. Wallis, Z. Allen-Zhu, Y. Li, S. Wang, L. Wang, W. Chen, et al., Lora: Low-rank adaptation of large language models., ICLR 1 (2) (2022) 3. 39

    2022

  51. [60]

    Bommasani, D

    R. Bommasani, D. A. Hudson, E. Adeli, R. Altman, S. Arora, S. von Arx, M.S.Bernstein, J.Bohg, A.Bosselut, E.Brunskill, E.Brynjolfsson, S. Buch, D. Card, R. Castellon, N. Chatterji, A. Chen, K. Creel, J. Q. Davis, D. Demszky, C. Donahue, M. Doumbouya, E. Durmus, S. Ermon, J. Et...

    2022 arXiv

  52. [61]

    Y. Bang, S. Cahyawijaya, N. Lee, W. Dai, D. Su, B. Wilie, H. Love- nia, Z. Ji, T. Yu, W. Chung, et al., A multitask, multilingual, multi- modal evaluation of chatgpt on reasoning, hallucination, and interactiv- ity, arXiv preprint arXiv:2302.04023 (2023)

    2023 arXiv

  53. [62]

    J. Lee, W. Yoon, S. Kim, D. Kim, S. Kim, C. H. So, J. Kang, Biobert: a pre-trained biomedical language representation model for biomedical text mining, Bioinformatics 36 (4) (2020) 1234–1240. 40

    2020

  54. [63]

    Chalkidis, M

    I. Chalkidis, M. Fergadiotis, P. Malakasiotis, N. Aletras, I. Androut- sopoulos, Legal-bert: The muppets straight out of law school, arXiv preprint arXiv:2010.02559 (2020)

    2020 arXiv

  55. [64]

    URL https://api.semanticscholar.org/CorpusID:268232499

    The claude 3 model family: Opus, sonnet, haiku. URL https://api.semanticscholar.org/CorpusID:268232499

  56. [65]

    URL https://openai.com/index/hello-gpt-4o/

    OpenAI, Hello gpt-4o, Online. URL https://openai.com/index/hello-gpt-4o/

  57. [67]

    D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al., Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning, arXiv preprint arXiv:2501.12948 (2025). 41

    2025 arXiv

  58. [517]

    doi:10.18653/v1/2023.findings-emnlp.36

    2023 doi

Pith tools

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