Pith. sign in

REVIEW 4 major objections 6 minor 40 references

Multi-Amateur Contrastive Decoding for Text Generation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read An ensemble of amateur models, rather than a single one, can steer a large language model toward more coherent and diverse text at inference time, with no retraining.

desk verdict A promising but incomplete extension of contrastive decoding: the mean-averaging variant is well-defined, but the consensus variant that drives the reported gains is not, making the paper non-reproducible as written. read the letter →

arxiv 2507.21086 v1 pith:WVWTOKFF submitted 2025-06-22 cs.CL

classification cs.CL
keywords multi-amateurcontrastivedecodingensembleopen-endedtextgenerationinference-timeoptimizationconsensuspenalizationcontrollablestrategies
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 MACD, an inference-time decoding strategy that generalizes contrastive decoding (CD) by replacing the single small amateur model with an ensemble of amateur models. The paper claims that this ensemble, combined through either mean or consensus aggregation, captures a wider range of generation failures such as repetition, hallucination, and stylistic drift, and that it consistently beats standard decoding baselines and original CD on diversity, coherence, fluency, and controllability. No training or fine-tuning is needed; only frozen pretrained models are used at inference. The strongest quantitative evidence reported is a 57.6 percent human preference for MACD over CD on coherence and a 63.2 percent preference over nucleus sampling on coherence, with fluency at parity. If this claim holds, the practical recipe for better text is not a bigger expert but a small committee of deliberately diverse weak critics.

What carries the argument

The load-bearing object is the ensemble contrastive score. Mean penalization computes $S_{\mathrm{mean}}(x)=\log P_E(x\mid x_{<t})-\alpha\frac{1}{K}\sum_k \log P_{A^{(k)}}(x\mid x_{<t})$, while consensus penalization first computes a consensus ratio $CR(x)=\frac{1}{K}\sum_k I_x^{(k)}$ over top-$r$ indicators and then scores $S_{\mathrm{consensus}}(x)=\log P_E(x\mid x_{<t})-\alpha\,C(x)$. The amateur ensemble is deliberately heterogeneous: compact models, domain-biased or distilled models, and n-gram pseudo-models. These scores sit on top of the CD plausibility filter, which keeps only tokens whose expert log-probability is within a margin $\delta$ of the mode. The whole procedure is zero-shot and inference-only.

What would settle it

Construct a prompt where the expert prefers a formal continuation but all selected amateurs were fine-tuned on the same informal corpus, so they agree on an informal token. If MACD's consensus penalty suppresses the expert's formal choice more than the mean penalty does, and human raters prefer the formal completion, then the consensus assumption is doing harm rather than good in that setting.

Watch

Extended reading notes

Core claim

The central claim is that penalizing tokens endorsed by a set of amateur language models, rather than by a single amateur, gives more coherent and diverse open-ended text while maintaining fluency. MACD defines two scoring rules: mean penalization, which subtracts the average amateur log-probability from the expert log-probability, and consensus penalization, which applies a threshold penalty proportional to how many amateurs rank the token in their top set. The plausibility filter is extended so that joint expert and amateur confidence determines which tokens survive. Experiments on WikiNews, WikiText-103, and BookCorpus report that MACD outperforms top-k, nucleus, typical, greedy, and original CD on diversity and coherence, and biases inserted into the amateur ensemble shift formality scores from 0.74 to 0.81-0.83 without hurting other metrics. The paper concludes that a small diverse ensemble of three amateurs is sufficient, with gains saturating beyond that.

Load-bearing premise

The consensus variant assumes that when several amateur models agree that a token is likely, that agreement is a reliable sign of degeneration; if the amateurs share a common bias instead, the consensus penalty could suppress perfectly valid continuations and the advantage over standard CD would depend on mean averaging alone.

Editorial extensions

If this is right

  • If MACD's coherence gains hold, open-ended generators can improve output quality by adding a small, diverse amateur ensemble at decoding time rather than retraining the expert.
  • Because the amateurs are interchangeable and frozen, the same expert can be steered toward different styles or domains by swapping the ensemble; adding an informal or toxic-biased amateur raises formality scores from 0.74 to 0.81-0.83.
  • The saturation at three amateurs implies that a small ensemble gives most of the benefit, and consensus aggregation costs only modest extra latency over mean aggregation.
  • The zero-shot property means the method can be applied to any frozen pretrained autoregressive model family, such as GPT-2 or OPT, without architectural change.

Reading between the lines

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

  • A testable extension is adaptive weighting: instead of equal mean weights, weight each amateur by its estimated reliability per domain, which might shift the saturation point and improve robustness.
  • If consensus agreement is indeed a degeneration signal, MACD could also serve as a diagnostic tool, ranking tokens by consensus ratio to expose clichés or repetitions shared across model families.
  • The framework could transfer to machine translation or code generation by choosing amateurs with known error biases, although the paper only demonstrates open-ended text.
  • One risk the reported experiments do not isolate is that consensus among amateurs may reflect shared pretraining biases rather than true degeneration; a control experiment with deliberately correlated amateurs would separate these effects.
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 / 6 minor

Summary. The paper proposes Multi-Amateur Contrastive Decoding (MACD), an inference-time extension of contrastive decoding in which an ensemble of small "amateur" language models provides the penalty signal for a large expert model. Two aggregation strategies are described: mean penalization, which averages amateur log-probabilities, and consensus penalization, which is intended to penalize tokens on which many amateurs agree. The authors also claim an ensemble-aware plausibility constraint and demonstrate results on WikiNews, Wikipedia, and story domains using automatic metrics, human evaluation, ablations, and runtime measurements, concluding that MACD consistently outperforms CD and other baselines without training.

Significance. If the method were fully specified and the results held, MACD would be a useful, low-cost generalization of CD: it is zero-shot, requires no fine-tuning, and the mean-penalization variant is simple and well defined. The paper also provides a reasonable breadth of evaluation across domains, a human preference study, an ablation, and a complexity analysis, which are appropriate steps for an empirical decoding paper. However, the current manuscript does not yet establish the central claim because the consensus mechanism, which the ablation identifies as critical, is not formally defined, and the statistical evidence is incomplete.

major comments (4)
  1. [IV.B, Eqs. (5)-(6)] The consensus penalization strategy is not defined. Eq. (6) defines S_consensus(x) = log P_E(x) - alpha * C(x), but C(x) is never introduced anywhere in the paper. Eq. (5) defines only CR(x), the consensus ratio. The prose and pseudocode in Section IV.C refer to a consensus threshold tau, but tau does not appear in Eq. (6). If C(x) is meant to be CR(x), then the penalty is alpha times a fraction in [0,1], which is not the described thresholded suppression; if C(x) is meant to be an indicator such as 1[CR(x) >= tau], the text must say this and specify tau and the top-r rank parameter. As written, the consensus variant cannot be implemented, reproduced, or tested, yet Table IV attributes the main gains of MACD to consensus penalization.
  2. [IV.A-IV.C] The promised ensemble-aware plausibility constraint is missing. The abstract and Section IV introduction state that MACD extends the plausibility constraint to the multi-amateur setting by introducing joint thresholds over expert confidence and amateur collective judgment, but Section IV.B merely says the candidate set is obtained using the Section III constraint, and no joint-threshold equation or algorithmic step appears. Either the constraint must be formally specified or the claim should be removed from the abstract and contributions.
  3. [V, Tables II-IV] The experimental reporting does not support the headline statistical claim. Table II reports a single "MACD" row per domain without stating whether mean or consensus aggregation was used; Table IV's ablation then attributes the improvement to consensus, whose definition is missing. The paper states "statistically significant improvements with p < 0.01 for paired t-test" but gives no sample sizes, number of prompts, variance estimates, seeds, or test details. Without these, a reader cannot verify that the 0.02-0.05 metric differences are not noise. Please report all numbers with standard errors or confidence intervals and define the test set and number of generations.
  4. [V, Models and hyperparameters] The claim of consistent superiority is weakened by the fact that alpha, amateur temperature, the number of amateurs, and the consensus threshold are tuned on the same evaluation domains. The paper states alpha=0.1 "found to generalize well" and tau settings for GPT-2 vs OPT, but no held-out validation or sensitivity analysis on unseen domains is provided. If these choices were selected after inspecting the reported tables, the comparison is partially circular. A sensitivity analysis on held-out domains or a clear description of how hyperparameters were chosen is needed.
minor comments (6)
  1. [IV.C] The pseudocode discussion says "In Fig 1, P_E(...) and P_Ak(...) denote..." but the described variables appear in Fig. 3, not Fig. 1; the figure reference should be corrected.
  2. [V, Table VI] Table VI is captioned as an "ABLATION STUDY" but it reports decoding times, not an ablation; the caption should be fixed to match the table content.
  3. [V, Table II] The headers "MAUVE-1," "Diversity-1," and "Coherence-1" use an unexplained suffix; it is unclear whether this denotes a particular run, iteration, or confidence level.
  4. [II, Related work] The sentence attributing "effective messages should be both relevant and not repetitive" to reference [7] is a citation mismatch: [7] is the ACL 2023 contrastive decoding paper, not a communication-theoretic source. Please cite the appropriate work or remove the attribution.
  5. [References] Several references contain typos, e.g., "transfomer" in [2] and "wizrd-of-Oz" in [15]; these should be corrected.
  6. [IV.C, time complexity] The complexity formulas in Eqs. (7)-(8) and Table I are not fully consistent: Table I defines CD as O(r*T) and MACD worst-case as O(K*r*T) with r undefined, while the text uses M and V. Align the notation among the text, equations, and table.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claim is supported by external empirical comparisons and explicit in-paper definitions, not by fitting or by a self-citation chain.

full rationale

The paper's central claim is an empirical one: MACD outperforms nucleus sampling, top-k sampling, typical sampling, greedy decoding, and original CD as measured by MAUVE, diversity, coherence, formality, and human preference. The method is defined in Sections IV.A-IV.C: Eq. (4) is an explicit K-amateur generalization of the single-amateur CD score in Eq. (2), and Eq. (5) defines the consensus ratio. These are design choices, not results derived from the same data on which they are later tested. The only self-citations are related-work entries [22]-[25] describing the authors' earlier decoding papers; they are contextual and carry no weight in the derivation of MACD's reported performance, so they are not load-bearing. The manuscript does have serious reproducibility defects that are not circularity: Eq. (6) refers to an undefined C(x), the consensus threshold tau introduced in the pseudocode never appears in the scoring rule, and alpha is set to 0.1 after being described as "found to generalize well" without a reported validation protocol. These are matters of correctness, completeness, and potential overfitting, not reductions of the reported predictions to the model's inputs by construction. Accordingly, no circular step meeting the quoted-equation evidentiary standard can be identified.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The method rests on the CD assumption plus several free hyperparameters. The mean formula is straightforward, but the consensus and plausibility-extension components are under-specified, so the reader must supply or guess implementation details.

free parameters (6)
  • alpha (contrastive penalty strength) = 0.1
    Set for all main experiments (Section V); selected as generalizing well across domains, so it is tuned on the evaluation domains.
  • amateur temperature tau = 0.5 (GPT-2), 1.0 (OPT)
    Chosen per model family in Section V; not justified by a separate validation set.
  • plausibility margin delta = not stated
    Inherited from CD equation (3) and used in MACD filtering; no value is given, so the candidate set size is unspecified.
  • consensus threshold tau (or C function parameters) = not stated
    Mentioned in Figure 3 caption as determining the penalty trigger; the equation for C(x) and the threshold value are missing.
  • top-r rank threshold for amateur indicators = not stated
    Equation (5) defines I_x^(k) as top-r, but r is unspecified.
  • number of amateurs K = 3
    Set to 3 in main experiments based on ablation saturation (Section V), a design choice made from the same evaluation data.
assumptions (4)
  • domain assumption Smaller amateur models exhibit the failure modes (repetition, incoherence, hallucination) that should be suppressed in the expert output.
    Inherited from CD (Section III) and asserted in Section IV.A; no new evidence is provided in this paper.
  • ad hoc to paper Consensus among amateur models on a token is a reliable signal of degeneration, while disagreement signals creativity.
    Introduced in Section IV.B as the basis for consensus penalization; not derived from theory or prior results.
  • domain assumption A small, diverse set of three amateurs spans the relevant failure modes.
    The ablation study in Section V shows diminishing returns after three amateurs, but this is an empirical claim on specific domains, not a general guarantee.
  • domain assumption The expert plausibility filter (Equation 3) remains appropriate when multiple amateurs are added.
    The paper claims an ensemble-aware adaptation but does not provide the modified constraint; it reuses the single-amateur filter.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Amateur Contrastive Decoding for Text Generation." pith.science (2026). https://pith.science/paper/WVWTOKFF

@misc{pith2026250721086,
  author       = {Pith},
  title        = {Pith review of: Multi-Amateur Contrastive Decoding for Text Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WVWTOKFF}},
  note         = {Machine review of arXiv:2507.21086}
}
read the original abstract

Contrastive Decoding (CD) has emerged as an effective inference-time strategy for enhancing open-ended text generation by exploiting the divergence in output probabilities between a large expert language model and a smaller amateur model. Although CD improves coherence and fluency, its dependence on a single amateur restricts its capacity to capture the diverse and multifaceted failure modes of language generation, such as repetition, hallucination, and stylistic drift. This paper proposes Multi-Amateur Contrastive Decoding (MACD), a generalization of the CD framework that employs an ensemble of amateur models to more comprehensively characterize undesirable generation patterns. MACD integrates contrastive signals through both averaging and consensus penalization mechanisms and extends the plausibility constraint to operate effectively in the multi-amateur setting. Furthermore, the framework enables controllable generation by incorporating amateurs with targeted stylistic or content biases. Experimental results across multiple domains, such as news, encyclopedic, and narrative, demonstrate that MACD consistently surpasses conventional decoding methods and the original CD approach in terms of fluency, coherence, diversity, and adaptability, all without requiring additional training or fine-tuning.

Figures

Figures reproduced from arXiv: 2507.21086 by the authors.

Figure 1
Figure 1. depicts the architecture of the MACD framework. At the heart of MACD lies the idea of ensemble-based penalization. Let 𝐴 = {𝐴 (1) ,𝐴 (2) , … 𝐴 (𝑘) } denote a collection of 𝐾 amateur models. Each amateur model 𝐴 (𝑘) defines a conditional distribution 𝑃𝐴 (𝑘) (𝑥𝑡 |𝑥<𝑡) over the next token 𝑥𝑡 , conditioned on the prefix 𝑥<𝑡 . These models may vary in architecture, pretraining data, size, or training objectives, thus nat… view at source ↗
Figure 2
Figure 2. depicts workflow of the MACD framework. The MACD procedure is fully zero-shot and inference-time only and does not need any training or fine-tuning of the expert or amateur models. The ensemble can be constructed offline and adjusted dynamically depending on the task or domain [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. The pseudocode for the MACD decoding procedure Time complexity for CD: In CD with a single amateur model, the expert model involves one forward pass yielding logits for all V tokens, and filtering (e.g., top-k) reduces this to M candidates. The amateur model requires M forward passes, one for each candidate 𝑥 ∈ 𝜈𝑡 to compute 𝑙𝑜𝑔𝑃𝐴(𝑥|𝑥<𝑡), and each pass requires a softmax computation and log-probability extraction fo… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: The trade-off between computation time and output quality While achieving high-quality generation is essential, for real-time or large-scale deployment, decoding methods must also be evaluated in terms of computational cost [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Fluency and repetition rate across decoding methods and domains Finally, several candidate decoding methods were evaluated based on fluency and repetition rates across different domains, with the results shown in [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 34 canonical work pages

  1. [1]

    Improving language understanding by generative pre -training,

    A. Radford, J. Wu, R. Child, D. Luan, D. Amodei, and I. Sutskever, “Improving language understanding by generative pre -training,” OpenAI Research Paper, 2018

  2. [2]

    OPT: Open pre -trained transfomer language models,

    S. Zhang, S. Roller, N. Goyal, M. Artetxe, M. Chen, S. Chen, C. Dewan, M. Diab, X. Li, X. V. Lin, T. Mihaylov, M. Ott, S. Shleifer, K. Shuster, D. Simig, P.S.Koura, A. Sridhar, T. Wang, and L. Zettlemoyer, “OPT: Open pre -trained transfomer language models," arXiv:2205.01068, 2022

  3. [3]

    Exploring the limits of transfer learning with a unified text-to-text transformer,

    C. Raffel, N. Shazeer, A. Roberts, K. Lee, S. Narang, M. Matena, Y. Zhou, W. Li, P. J. Liu, “Exploring the limits of transfer learning with a unified text-to-text transformer,” Journal of Machine Learning Research, Vol 21, No 140, 1–67, 2019

  4. [4]

    Hierarchical neural story generation,

    A. Fan, M. Lewis, and Y. Dauphin, “Hierarchical neural story generation,” arXiv:1805.0483, 2018

  5. [5]

    The curious case of neural text degeneration,

    A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi, “The curious case of neural text degeneration,” in: Proc. of the 8 th Int. Conf. on Learning Representations (ICLR), 2020

  6. [6]

    Locally typical sampling,

    C. Meister, T. Pimentel, G. Wiher, R. Cotterell, “Locally typical sampling,” Transactions of the Association for Computational Linguistics, Vol 11, pp. 102-121, MIT Press, MA, USA, 2023

  7. [7]

    Contrastive decoding: Open-ended text generation as optimization,

    X.L. Li, A. Holtzman, D. Fried, P. Liang, J. Eisner, T. Hashimoto, L. Zettlemoyer, and M. Lewis, “Contrastive decoding: Open-ended text generation as optimization,” in: Proc. of the 61st Annual Meeting of the ACL, Vol 1, pp. 12286-12312, 2023

  8. [8]

    MAUVE Scores for Generative Models: Theory and Practice,

    K. Pillutla , L. Liu, J . Thickstun, S . Welleck, S . Swayamdipta, R . Zellers, S . Oh, Y . Choi, and Z. Harchaoui, “ MAUVE Scores for Generative Models: Theory and Practice, ” Journal of Machine Learning Research (JMLR), Vol 24, No 1-92, 2022

Show all 40 references
  1. [9]

    A contrastive framework for neural text generation,

    Y. Su, T. Lan, Y. Wang, D. Yogatama, L. Kong, and N. Co llier, “A contrastive framework for neural text generation,” in: Proc. of the 36th NIPS’22, pp. 21548-21561, 2022

  2. [10]

    Adaptive contrastive search: Uncertainty -guided decoding for open -ended text generation,

    E. Garces Arias, J. Rodemann, M. Li, C. Heumann, and M. Aßenmacher, “Adaptive contrastive search: Uncertainty -guided decoding for open -ended text generation,” Findings of the ACL: EMNLP’24, pp. 15060-15080, 2024

  3. [11]

    CoNT: Contrastive neural text generation,

    C. An, J. Feng, K. Lv, L. Kong, X. Qiu, and X. Huang , “CoNT: Contrastive neural text generation,” in: Proc. of the 36th NeurIPS, 2022

  4. [12]

    Surfacing biases in large language models using contrastive input decoding,

    G. Yona, O. Honovich, I. Laish, and R. Aharoni, “Surfacing biases in large language models using contrastive input decoding,” arXiv:2305.07378, 2023

  5. [13]

    On decoding strategies for neural text generators,

    G. Wiher, C. Meister, and R. Co tterell, “On decoding strategies for neural text generators,” TACL, Vol 10, pp. 997-1012, 2022

  6. [14]

    DecoStrat: Leveraging the capabilities of language models in D2T generation via decoding framework,

    E. L. Jimale, W. Chen, M. A. Al -antari, Y. H. Gu, V. K. Agbesi, and W. Feroze, “DecoStrat: Leveraging the capabilities of language models in D2T generation via decoding framework,” Mathematics, Vol 12, No 22, Art ID: 3596, 2024

  7. [15]

    MultiWOZ- A large-scale multi-domain wizrd- of-Oz dataset for task-oriented dialogue modelling,

    P. Budzianowski, T -H. Wen, B-H. Tseng, I. Casanueva, S. U ltes, O. Ramadan, M. Gašić, “MultiWOZ- A large-scale multi-domain wizrd- of-Oz dataset for task-oriented dialogue modelling,” ENLP, 2018

  8. [16]

    Speculative contrastive decoding,

    H. Yuan, K. Lu, F. Huang, Z. Yuan, a nd C. Zhou, “Speculative contrastive decoding,” arXiv: 2311.08981, 2023

  9. [17]

    Knowledge infused decoding,

    R. Liu , G. Zheng, S . Gupta, R . Gaonkar, C . Gao, S . Vosoughi, M . Shokouhi, A. H. Awadallah, “Knowledge infused decoding,” in: Proc. of ICLR’22, 2022

  10. [18]

    Improving machine translation with large language models: A preliminary study with cooperative decoding,

    J. Zeng, F. Meng, Y. Yin, and J. Zhou, “Improving machine translation with large language models: A preliminary study with cooperative decoding,” Findings of the ACL, pp. 13275-13288, 2024

  11. [19]

    C. Wang, J. Zhang, and H. Chen, “Semi-autoregressive neural machine translation,“ in: Proc. of the Conf on EMNLP, pp. 479-488, 2018

  12. [20]

    Mutual information and diverse decoding improve neural machine translation,

    J. Li and D. Jurafsky, “Mutual information and diverse decoding improve neural machine translation,” arXiv: 1601.00372, 2016

  13. [21]

    Diverse beam search for improved description of complex scenes,

    A. Vijayakumar, M. Cogswell, R. Selvaraju, Q. Sun, S. Lee, D. Crandall, and D. Batra, “Diverse beam search for improved description of complex scenes,” in: Proc. of the AAAI Conf. on Artificial Intelligence, Vol 32, No 1, 2018

  14. [22]

    Contextual contrastive search for improved text generations in large language model,

    J. Sen, H. Waghela, and R.Pandey, “Contextual contrastive search for improved text generations in large language model,” in: Proc of the 6th IEEE Conf on Emerging Technologies (INCET’25), May 22-24, 2025, Belgaum, India (In press)

  15. [23]

    Advancing decoding strategies: Enhancements in locally typical sampling for LLMs,

    J. Sen, S. Sengupta, and Subhasis Dasgupta, “Advancing decoding strategies: Enhancements in locally typical sampling for LLMs,” in: W. Pedrycz, G. Rivera, R.M. Rodriguez, and S. Ibarra (eds). Decision Making in Computational Intelligence -Based Systems , Springer Nature, Switz...

  16. [24]

    Confidence -Modulated Speculative Decoding for Large Language Models,

    J. Sen, S. Dasgupta, and H. Waghela, “Confidence -Modulated Speculative Decoding for Large Language Models,” in: Proc of 6 th IEEE Indian Council International Subsections Conference (IEEE INDISCON 2025), August 21-23, 2025, Rourkela, India. (Accepted for publication)

  17. [25]

    Hierarchical Verification of Speculative Beams for Accelerating LLM Inference,

    J. Sen, H. Puvvala, and S. Dasgupta, “Hierarchical Verification of Speculative Beams for Accelerating LLM Inference,” in: Proc of 3rd Int Conf on Data Science and Network Engineering (ICDNE 2025) , July 25-26, 2025, Agartala, India. (Accepted for publication)

  18. [26]

    Energy -based reranking: Improving neural machine translation using energy -based models,

    S. Bhattacharyya, A. Rooshenas, S. Naskar, S. Sun, M. Iyyer, and A. McCallum, “Energy -based reranking: Improving neural machine translation using energy -based models,” in: Proc. of the 59 th Annual Meeting of the ACL and the 11th Int Joint Conf on NLP, Vol 1, pp. 4528- 4537, 2021

  19. [27]

    Plug and play language models: A simple approach to controlled text generation,

    S. Dathathri, A. Madotto, J. Lan, J. Hung, E. Frank, P. Molino , J. Yosinski, and R. Liu, “Plug and play language models: A simple approach to controlled text generation,” in: Proc. of ICLR 2020

  20. [28]

    Entropy- regularized to ken-level policy optimization for language agent reinforcement,

    M. Wen, J. Liao, C. Deng, J. Wang, W. Zhang, and Y. Wen, “Entropy- regularized to ken-level policy optimization for language agent reinforcement,” arXiv:2402.06700, 2024

  21. [29]

    Ensemble methods in machine learning,

    T. G. Dietterich, “Ensemble methods in machine learning,” in: Proc. of Int Workshop on Multiple Classifier Systems, pp. 1-15, Springer, 2000

  22. [30]

    Url: https://www.wikinews.org/

    WikiNews. Url: https://www.wikinews.org/

  23. [31]

    Pointer sentinel mixture models,

    S. Merity, C. Xiong, J. Bradbury, and R. Socher, “Pointer sentinel mixture models,” arXiv:1609.07843, 2016

  24. [32]

    Aligning books and movies: Towards story -like visual explanations by watching movies and reading books,

    Y. Zhu, R. Kiros, R. Zemel, R. Salakhutdinov, R. Urtasun, A. Torralba, and S. Fidler, “Aligning books and movies: Towards story -like visual explanations by watching movies and reading books,” arXiv:1506.06724, 2015

  25. [33]

    SimCSE: Simple contrastive learning of sentence embeddings,

    T. Gao, X. Yao, and D. Chen, “SimCSE: Simple contrastive learning of sentence embeddings,” in: Proc. of the 2021 Conf. on EMNLP , pp. 6894-6910, 2021

  26. [34]

    An empirical analysis of formality in online communication,

    E. Pavlick and J. Tetreault, “An empirical analysis of formality in online communication,” Trans. of the ACL, Vol 4, pp. 61-74, 2016

  27. [35]

    Url: https://huggingface.co/openai-community/gpt2-xl

    GPT-2 XL. Url: https://huggingface.co/openai-community/gpt2-xl

  28. [36]

    Url: https://huggingface.co/facebook/opt-6.7b

    OPT-6.7B. Url: https://huggingface.co/facebook/opt-6.7b

  29. [37]

    Url: https://huggingface.co/facebook/opt-13b

    OPT-13B. Url: https://huggingface.co/facebook/opt-13b

  30. [38]

    Url: https://huggingface.co/huseinzol05/gpt2-117M

    GPT-2-117M. Url: https://huggingface.co/huseinzol05/gpt2-117M

  31. [39]

    Url: https://huggingface.co/facebook/opt-125m

    OPT-125M. Url: https://huggingface.co/facebook/opt-125m

  32. [40]

    Url: https://www.mturk.com/

    Amazon Mechanical Turk. Url: https://www.mturk.com/

Pith tools

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