Pith. sign in

REVIEW 4 major objections 7 minor 41 references

Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Input-dependent soft prompts generated by self-attention beat static prompt baselines across GLUE and SuperGLUE.

desk verdict A clean, honest soft-prompting paper whose main comparison set excludes the two closest input-dependent methods, so the SOTA claim needs work before it can be believed. read the letter →

arxiv 2506.05629 v1 pith:ZKZ55OMH submitted 2025-06-05 cs.CL

Ananth Muppidi , Abhilash Nandy , Sambaran Bandyopadhyay This is my paper
classification cs.CL
keywords input-dependentsoftpromptingself-attentionparameter-efficientfine-tuningpromptzero-shottransferGLUESupertuning
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 introduces ID-SPAM, a way to create soft prompts for frozen large language models that are generated from the input itself rather than being a fixed learned vector. The prompt is produced by a self-attention layer over the input token embeddings, followed by mean pooling and a two-layer bottleneck MLP, then prepended to the input of a single chosen transformer layer. The authors claim that this input-dependent prompt outperforms all parameter-efficient soft-prompt baselines on average accuracy across six GLUE tasks and four SuperGLUE tasks, and improves zero-shot task and domain transfer. The method keeps the number of trainable parameters small, comparable to late prompt tuning and below LoRA. An ablation attributes most of the gain to the self-attention layer, which adds about 5.8 percent average accuracy over mean pooling alone.

What carries the argument

The generator is a small trainable network: a single self-attention layer applied to the input token embeddings, with the attention output mean-pooled across tokens; the pooled vector passes through a down-projection MLP, ReLU, and an up-projection MLP, and the result is resized into the soft prompt $S_T \in \mathbb{R}^{n \times t}$, with $t=10$ prompt tokens. This prompt is prepended to the input of one transformer layer, whose index $m$ is a per-task hyperparameter. The self-attention supplies the input-dependence: different input tokens are weighted differently when the prompt is built. The bottleneck MLP ($c < n$) keeps the trainable parameter count small.

What would settle it

Permute the ordering of input tokens before generating the soft prompt while keeping the rest of the pipeline identical. If accuracy stays the same under token permutation, the self-attention is not using token identity or order, so the claimed input-dependence is not doing the work. Alternatively, a broader benchmark where ID-SPAM's average gain over late prompt tuning shrinks or reverses would falsify the generalization claim.

Watch

Extended reading notes

Core claim

The central claim is that conditioning a soft prompt on the actual input tokens, with the prompt generator attending to different tokens with different weights, makes parameter-efficient fine-tuning more accurate and more transferable than static soft prompts. Previous soft prompting methods learn one prompt vector per task and leave it fixed for every instance, whereas ID-SPAM computes a new prompt for each input. The paper shows this on six GLUE tasks and four SuperGLUE tasks with RoBERTa-BASE and RoBERTa-LARGE backbones: ID-SPAM outperforms all parameter-efficient soft-prompt baselines on 4 of 6 GLUE tasks and on average, gives the best mean score on SuperGLUE, and beats the soft-prompt baselines on zero-shot transfer between QQP/MRPC and SST-2/IMDB. It also reports better or comparable results than LoRA on average GLUE accuracy while training fewer parameters.

Load-bearing premise

The method assumes that a single trainable self-attention layer over input embeddings, mean-pooled and passed through a two-layer MLP, produces soft prompts that generalize across tasks and domains when prepended at a fixed per-task layer index $m$; the layer index is chosen by manual tuning, not automatically.

Editorial extensions

If this is right

  • If the central claim holds, frozen large language models can be adapted to a new task by training only a small prompt generator plus prompt tokens, achieving stronger accuracy than static soft prompting.
  • The improved zero-shot transfer (e.g., QQP to MRPC, SST-2 to IMDB) suggests that input-dependent prompts carry instance-level features that generalize across tasks and domains, not just task-specific biases.
  • Because the parameter savings over LoRA grow with backbone size, the method is well positioned for very large open models where adapter matrices become costly.
  • The layer index $m$ is a load-bearing hyperparameter; performance is best in middle layers and degrades at very early or late insertion points, so automated layer selection is a natural next step.
  • The method is backbone-agnostic: reported gains appear with both encoder (RoBERTa) and decoder-only (GPT-2, GPT-2 Large) architectures.

Reading between the lines

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

  • If the self-attention prompt generator is truly learning token importance, its attention weights could be visualized to identify which words drive task decisions, offering an interpretability tool for frozen LLMs.
  • A stress test beyond GLUE/SuperGLUE—e.g., named-entity recognition or question answering with long inputs—would show whether the single-layer mean-pooled prompt saturates as input length grows.
  • The paper's fixed layer index might interact with the prompt: the same generator could be connected to a learned router that selects the insertion layer per input, combining the method with adaptive-depth architectures.
  • Since the gain over mean pooling is large (about 5.8 percent), the self-attention component is the likely source of the improvement; ablating attention head count or replacing softmax with linear attention would test whether the effect is specifically attention-like weighting.
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 / 7 minor

Summary. The paper proposes ID-SPAM, a parameter-efficient fine-tuning method that generates an input-dependent soft prompt from the input token embeddings using a single self-attention layer followed by a two-layer MLP, and prepends the prompt at a chosen transformer layer of a frozen language model. The method is evaluated on six GLUE tasks and four SuperGLUE tasks with RoBERTa-BASE/LARGE and GPT-2 backbones, and on zero-shot task/domain transfer between four dataset pairs. The reported results indicate that ID-SPAM achieves higher mean accuracy than the implemented soft-prompt baselines on a majority of GLUE tasks and on average, and improves zero-shot transfer on the four evaluated pairs.

Significance. If the central claim holds, ID-SPAM is a simple and parameter-efficient alternative to existing soft-prompt methods, with the useful property that the prompt is generated from the input at inference time. The paper reports means and standard deviations over three runs for the main GLUE/SuperGLUE tables, and it includes an ablation showing the self-attention layer contributes beyond mean pooling. The method's parameter counts are modest and reported in the appendix. However, the comparison set excludes the most closely related input-dependent soft-prompt baselines, and the hyperparameters for baselines are not disclosed, so the claimed advantage over 'all parameter-efficient soft-prompt baselines' is not currently verifiable.

major comments (4)
  1. [Section 3.1-3.2, Tables 2, 4-6] The baseline set used to support the claim that ID-SPAM 'outperforms all Parameter-Efficient Soft Prompt-based baselines' excludes the very methods the introduction identifies as closest: IDPG (Wu et al., 2022) and APrompt (Wang et al., 2023), both of which generate input-dependent soft prompts. Since the paper's contribution is specifically input-dependent soft prompting, the reported advantage over Prompt Tuning, P-Tuning, SMoP, LPT, and DEPT could be an artifact of baseline selection. Please add IDPG and APrompt (and, if available, the attention-based instance-aware prompting of Jin et al., 2023) under the same experimental settings, or explicitly restrict the SOTA claim to the evaluated baselines.
  2. [Appendix A] Hyperparameter tuning is reported only for ID-SPAM (Table 7); for the baselines the text states 'we used the best hyperparameter set for each of the baselines' without reporting those sets or the search procedure. This omission makes it impossible to assess whether the comparison is fair, especially for LPT (layer index, prompt length) and DEPT (rank), which are known to be sensitive to such choices. Please report the tuned hyperparameters for every baseline and method.
  3. [Section 3.2, Table 2] The claim of 'outperforming' on 4 out of 6 GLUE tasks rests on point estimates whose standard deviations overlap substantially; for example, on RoBERTa-LARGE SST-2, ID-SPAM (94.6 ± 1.2) is lower than both P-Tuning and SMoP (95.8 ± 0.8 and 95.8 ± 1.4), and on MNLI the difference from SMoP is 87.4 ± 0.8 vs. 86.7 ± 1.1. With only three runs, the reported differences are not sufficient to establish statistical significance. Please provide significance tests or paired comparisons, and phrase the conclusion as 'higher mean accuracy on X of 6 tasks' where the differences are not significant.
  4. [Section 3.4, Table 6, and Section 5] The zero-shot transfer evaluation covers only four transfer pairs (QQP↔MRPC and SST-2↔IMDB), all within closely related task types (paraphrase detection and sentiment classification). The conclusion that ID-SPAM 'significantly improves ... zero-shot task and domain transfer performance on various NLU tasks' is stronger than what these four pairs support. Please either expand the transfer evaluation to more diverse task/domain pairs or temper the conclusion to the evaluated settings.
minor comments (7)
  1. [Section 2, Eq. (2)] Equation (2) writes 'b_down' twice; the second bias should presumably be b_up. Please correct.
  2. [Section 2, first paragraph] There is a duplicated word: 'for for a task with a single sentence' should read 'for a task with a single sentence'.
  3. [Section 1, introduction] The word 'diving' in 'diving the soft prompt into query, key and value prompts' should be 'dividing'.
  4. [Table 1] The column headers 'Labels' and 'Type' are unclear: 'Labels' appears to list the number of classes, and 'Type' lists the category. Please rename for clarity (e.g., '#Classes' and 'Category').
  5. [Table 2 and Tables 4-6] The caption states that the best performing soft-prompt-based method's results are bold, but some entries where the best is a baseline (e.g., SST-2 with RoBERTa-LARGE) appear inconsistently formatted. Please ensure the bold formatting matches the caption exactly.
  6. [Appendix D, Table 12] For RoBERTa-LARGE, ID-SPAM has 3,538,944 trainable parameters versus LPT's 2,883,584; the text calls this 'comparable,' which is fair, but a reader may appreciate an explicit note that ID-SPAM is slightly larger in this configuration.
  7. [Appendix E] The convergence table for LoRA on MRPC and RTE is useful, but the paper does not provide analogous convergence evidence for the other baselines; if the intent is to demonstrate that all baselines were trained to convergence, please clarify how this was verified for the remaining methods.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the method is an empirical architecture trained on labeled data and evaluated on held-out tasks and zero-shot transfer; reported claims do not reduce to fitted inputs or self-cited premises.

full rationale

The paper contains no derivation chain in which a prediction reduces to a fitted input or to a self-citation. Equations (1) and (2) define the soft prompt as a deterministic function of input embeddings with trainable attention and MLP weights; these weights are optimized on the training split and the reported results are measured on held-out test sets (Tables 2, 4, 5, 6). The zero-shot transfer protocol, where a model trained on QQP is evaluated on MRPC and a model trained on SST-2 is evaluated on IMDB, is an out-of-distribution evaluation and is not determined by construction. The layer index m is explicitly acknowledged as a hyperparameter without automated selection in the Limitations section, so it is not a fitted parameter renamed as a prediction. The ablation in Table 3 compares ID-SPAM to mean-pooling and shows a performance difference, which further indicates that the attention mechanism carries empirical content rather than being definitionally equivalent to the baseline. The skeptic's concern about omitted input-dependent baselines (IDPG, APrompt) is a legitimate empirical completeness concern about the comparison class, but omitting a baseline does not make the reported average accuracy equal to the method's own inputs by construction. No uniqueness theorem, no ansatz smuggled via self-citation, and no renaming of a known result are present. Therefore the correct circularity score is 0.

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

No invented entities. The free parameters are standard hyperparameters and learned weights. The method relies on established transformer attention as an axiom.

free parameters (3)
  • Soft prompt length t = 10
    Number of soft prompt tokens, fixed to 10 for all methods. Chosen by hand; affects the prompt capacity.
  • Layer index m = per-task between 1-23 (RoBERTa-Large) and 1-11 (RoBERTa-Base)
    Chosen via hyperparameter search per task; the paper notes there is no automated way to choose it (Limitations section).
  • Bottleneck dimension c = not reported
    The MLP bottleneck width c < n is not reported in the paper, though it determines the number of trainable parameters.
assumptions (3)
  • standard math Standard transformer attention (Vaswani et al. 2017) is a valid mechanism for mixing input token information.
    Used in Eq. (1) without justification beyond citation.
  • domain assumption Soft prompts prepended to a single transformer layer can effectively adapt a frozen LM.
    Assumed throughout; supported by prior work (Lester et al., LPT) but not proven for this generator.
  • domain assumption The mean-pooled attention output retains sufficient information to generate a task-relevant prompt.
    The design choice of mean over tokens is not justified; the ablation shows it beats mean-pooling but does not explain why.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs." pith.science (2026). https://pith.science/paper/ZKZ55OMH

@misc{pith2026250605629,
  author       = {Pith},
  title        = {Pith review of: Leveraging Self-Attention for Input-Dependent Soft Prompting in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKZ55OMH}},
  note         = {Machine review of arXiv:2506.05629}
}
read the original abstract

The performance of large language models in domain-specific tasks necessitates fine-tuning, which is computationally expensive and technically challenging. This paper focuses on parameter-efficient fine-tuning using soft prompting, a promising approach that adapts pre-trained models to downstream tasks by learning a small set of parameters. We propose a novel Input Dependent Soft Prompting technique with a self-Attention Mechanism (ID-SPAM) that generates soft prompts based on the input tokens and attends different tokens with varying importance. Our method is simple and efficient, keeping the number of trainable parameters small. We show the merits of the proposed approach compared to state-of-the-art techniques on various tasks and show the improved zero shot domain transfer capability.

Figures

Figures reproduced from arXiv: 2506.05629 by the authors.

Figure 1
Figure 1. ID-SPAM Framework. Given an LM, the generated soft-prompt can be prepended to any transformer layer’s inputs (the figure can be best seen in color) intermediate layer (Liu et al., 2022a), or transform the soft prompt by using cross-attention with the in￾put tokens without explicitly generating from them (Jin et al., 2023). These input dependent prompting techniques still have multiple limitations: (i) Many of them e… view at source ↗
Figure 2
Figure 2. Effect of Variation in layer index (m) corre￾sponding to which soft prompt is prepended on perfor￾mance (m = 0 refers to input embeddings). Metrics are average of acc. and F1 for MRPC and acc. for RTE. We analyze the effect of varying layer index where soft prompt is prepended (m in [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 18 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  4. [4]

    Joon-Young Choi, Junho Kim, Jun-Hyung Park, Wing-Lam Mok, and SangKeun Lee. 2023. Smop: Towards efficient and effective prompt tuning with sparse mixture-of-prompts. In The 2023 Conference on Empirical Methods in Natural Language Processing

    work page 2023

  5. [5]

    Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Yunxuan Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, et al. 2024. Scaling instruction-finetuned language models. Journal of Machine Learning Research, 25(70):1--53

    2024

  6. [6]

    Ido Dagan, Oren Glickman, and Bernardo Magnini. 2005. The pascal recognising textual entailment challenge. In Proceedings of the PASCAL Challenges Workshop on Recognising Textual Entailment, volume 1

    work page 2005

  7. [7]

    Damai Dai, Yutao Sun, Li Dong, Yaru Hao, Shuming Ma, Zhifang Sui, and Furu Wei. 2023. Why can gpt learn in-context? language models secretly perform gradient descent as meta-optimizers. In Findings of the Association for Computational Linguistics: ACL 2023, pages 4005--4019

    work page 2023

  8. [8]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. 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), pages 4171--4186

    2019

Show all 41 references
  1. [9]

    William B Dolan and Chris Brockett. 2005. Automatically constructing a corpus of sentential paraphrases. In Proceedings of the Third International Workshop on Paraphrasing (IWP)

    2005

  2. [10]

    Zeyu Han, Chao Gao, Jinyang Liu, Sai Qian Zhang, et al. 2024. Parameter-efficient fine-tuning for large models: A comprehensive survey. arXiv preprint arXiv:2403.14608

    2024 arXiv

  3. [11]

    Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2022. https://openreview.net/forum?id=nZeVKeeFYf9 Lo RA : Low-rank adaptation of large language models . In International Conference on Learning Representations

    2022

  4. [12]

    Feihu Jin, Jinliang Lu, Jiajun Zhang, and Chengqing Zong. 2023. Instance-aware prompt learning for language understanding and generation. ACM Transactions on Asian and Low-Resource Language Information Processing, 22(7):1--18

    2023

  5. [13]

    Brian Lester, Rami Al-Rfou, and Noah Constant. 2021. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045--3059

    2021

  6. [14]

    Rush, and Thomas Wolf

    Quentin Lhoest, Albert Villanova del Moral, Patrick von Platen, Suraj Patil, Julien Chaumond, Mariama Drame, Timo M \"u ller, Isabella G \'e ron, Simon Brandeis, Sylvain Gugger, Th \'e o Matussi \`e re, Abhishek Thakur, Philipp Schmid, Yacine Jernite, Jeff Boudier, Francesco C...

    2021 arXiv

  7. [15]

    Xiang Lisa Li and Percy Liang. 2021. Prefix-tuning: Optimizing continuous prompts for generation. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Lo...

    2021

  8. [16]

    Xiangyang Liu, Tianxiang Sun, Xuan-Jing Huang, and Xipeng Qiu. 2022 a . Late prompt tuning: A late prompt could be better than many prompts. In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 1325--1338

    2022

  9. [17]

    Xiao Liu, Kaixuan Ji, Yicheng Fu, Weng Tam, Zhengxiao Du, Zhilin Yang, and Jie Tang. 2022 b . P-tuning: Prompt tuning can be comparable to fine-tuning across scales and tasks. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 2:...

    2022

  10. [18]

    Xiao Liu, Yanan Zheng, Zhengxiao Du, Ming Ding, Yujie Qian, Zhilin Yang, and Jie Tang. 2021. Gpt understands, too. arXiv:2103.10385

    2021 arXiv

  11. [19]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. https://arxiv.org/abs/1907.11692 Roberta: A robustly optimized bert pretraining approach . Preprint, arXiv:1907.11692

    2019 arXiv

  12. [20]

    Ilya Loshchilov and Frank Hutter. 2018. Decoupled weight decay regularization. In International Conference on Learning Representations

    2018

  13. [21]

    Fang Ma, Chen Zhang, Lei Ren, Jingang Wang, Qifan Wang, Wei Wu, Xiaojun Quan, and Dawei Song. 2022. Xprompt: Exploring the extreme of prompt tuning. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11033--11047

    2022

  14. [22]

    Vinod Nair and Geoffrey E. Hinton. 2010. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th International Conference on International Conference on Machine Learning, ICML'10, page 807–814, Madison, WI, USA. Omnipress

    2010

  15. [23]

    Aleksandar Petrov, Philip Torr, and Adel Bibi. 2023. When do prompting and prefix-tuning work? a theory of capabilities and limitations. In The Twelfth International Conference on Learning Representations

    2023

  16. [24]

    Guanghui Qin and Jason Eisner. 2021. https://doi.org/10.18653/v1/2021.naacl-main.410 Learning how to ask: Querying LM s with mixtures of soft prompts . In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human L...

    2021 doi

  17. [25]

    Quora. 2017. https://www.quora.com/q/quoradata/First-Quora-Dataset-Release-Question-Pairs Quora question pairs

    2017

  18. [26]

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Percy Liang. 2016. Squad: 100,000+ questions for machine comprehension of text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 2383--2392

    2016

  19. [27]

    Zhengxiang Shi and Aldo Lipani. 2024. https://openreview.net/forum?id=KjegfPGRde De PT : Decomposed prompt tuning for parameter-efficient fine-tuning . In The Twelfth International Conference on Learning Representations

    2024

  20. [28]

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. 2013. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language...

    2013

  21. [29]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya Pathak, Laurent Sifre, Morgane Rivi \`e re, Mihir Sanjay Kale, Juliette Love, et al. 2024. Gemma: Open models based on gemini research and technology. arXiv preprint arXiv:2403.08295

    2024 arXiv

  22. [30]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf Attention is all you need . In Advances in Ne...

    2017

  23. [31]

    Alex Wang, Yada Pruksachatkun, Nikita Nangia, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2019. Superglue: A stickier benchmark for general-purpose language understanding systems. Advances in neural information processing systems, 32

    2019

  24. [32]

    Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel Bowman. 2018. https://doi.org/10.18653/v1/W18-5446 GLUE : A multi-task benchmark and analysis platform for natural language understanding . In Proceedings of the 2018 EMNLP Workshop B lackbox NLP : A...

    2018 doi

  25. [33]

    Qifan Wang, Yuning Mao, Jingang Wang, Hanchao Yu, Shaoliang Nie, Sinong Wang, Fuli Feng, Lifu Huang, Xiaojun Quan, Zenglin Xu, et al. 2023. Aprompt: Attention prompt tuning for efficient adaptation of pre-trained language models. In Proceedings of the 2023 Conference on Empiri...

    2023

  26. [34]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in Neural Information Processing Systems, 35:24824--24837

    2022

  27. [35]

    Adina Williams, Nikita Nangia, and Samuel Bowman. 2018. A broad-coverage challenge corpus for sentence understanding through inference. In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technolo...

    2018

  28. [36]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R \'e mi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, ...

    2020

  29. [37]

    Zhuofeng Wu, Sinong Wang, Jiatao Gu, Rui Hou, Yuxiao Dong, VG Vinod Vydiswaran, and Hao Ma. 2022. Idpg: An instance-dependent prompt generation method. In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human L...

    2022

  30. [38]

    Yee Hui Yeo, Jamil S Samaan, Wee Han Ng, Peng-Sheng Ting, Hirsh Trivedi, Aarshi Vipani, Walid Ayoub, Ju Dong Yang, Omer Liran, Brennan Spiegel, et al. 2023. Assessing the performance of chatgpt in answering questions regarding cirrhosis and hepatocellular carcinoma. medRxiv, p...

    2023

  31. [39]

    Haopeng Zhang, Xiao Liu, and Jiawei Zhang. 2023 a . Summit: Iterative text summarization via chatgpt. arXiv preprint arXiv:2305.14835

    2023 arXiv

  32. [40]

    Zhen-Ru Zhang, Chuanqi Tan, Haiyang Xu, Chengyu Wang, Jun Huang, and Songfang Huang. 2023 b . Towards adaptive prefix tuning for parameter-efficient language model fine-tuning. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2...

    2023

  33. [41]

    Wei Zhu and Ming Tan. 2023. Spt: Learning to selectively insert prompts for better prompt tuning. In The 2023 Conference on Empirical Methods in Natural Language Processing

    2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.