Pith. sign in

REVIEW 3 major objections 5 minor 68 references

Writing Like the Best: Exemplar-Based Expository Text Generation

T0 review · 3 major / 5 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 RePA generates a factual new-topic text from one similar exemplar, beating prompting and retrieval-augmented baselines on factuality and style metrics.

desk verdict A genuinely new task plus a plausible prompting framework; the empirical case is solid but missing significance tests and calibration analysis for the refusal threshold. read the letter →

arxiv 2505.18859 v1 pith:VMJ7K5ZP submitted 2025-05-24 cs.CL cs.AI

Yuxiang Liu , Kevin Chen-Chuan Chang This is my paper
classification cs.CLcs.AI
keywords exemplar-basedtextgenerationexpositoryplan-then-adaptrecurrentpromptingconfidencecalibrationretrieval-augmentedLLM-as-a-judgelong-form
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 Exemplar-Based Expository Text Generation: given one well-written expository text about a source topic, produce a comparable text about a different but similar topic, preserving the exemplar's structure while replacing its facts with correct ones for the target topic. To do this it proposes RePA, a Recurrent Plan-then-Adapt framework that processes the exemplar sentence by sentence, turns each clarified segment into questions, transfers those questions to the target topic by topic-token substitution, and answers them with retrieval-augmented generation plus a confidence-based refusal that drops answers the LLM is unsure about. The paper claims RePA outperforms the listed baselines—direct prompting, rolling prompting, o1, self-refine, and retrieval-augmented versions of each—on factuality and on new LLM-judged metrics of imitativeness and adaptiveness across three collected datasets. If the claim holds, RePA offers a practical one-shot recipe for high-volume expository writing tasks such as university overviews, event introductions, and product descriptions, with less hallucination than direct prompting.

What carries the argument

The load-bearing mechanism is RePA's recurrent unit, which mirrors the recurrence of an LSTM but uses text in place of vectors and an LLM in place of the cell computation. Clarify uses short-term memory $h_t$ to resolve pronouns in the incoming segment; Outline prompts the LLM for questions about the source topic and then substitutes source-topic tokens with target-topic tokens, making the exemplar's talking points transferable plans. Calibrated-QA retrieves per-topic and per-query knowledge and answers each question while emitting a verbalized confidence score, filtering out answers below $\theta=0.7$; Write drafts the target segment from the surviving adapted facts and revises it against long-term memory $c_t$, a running summary of prior output segments. Questions-as-outlines carry the imitation across topics, and the calibrated refusal is what keeps the adapted content factual.

What would settle it

Run RePA on the Wikipedia or USNews test pairs with the Calibrated-QA confidence threshold swept from 0 to 1 while keeping everything else fixed. If lowering the threshold does not increase hallucination or raising it does not improve NLI entailment per output word, then the verbalized-confidence refusal is not the active cause of the factuality gain. A complementary check: measure the correlation between the LLM's verbalized confidence on individual answers and whether that answer is entailed by the ground truth; if confident-but-wrong answers are common, the refusal gate is miscalibrated for this task.

Watch

Extended reading notes

Core claim

The central claim is that a long-form expository text on a new topic can be produced from a single exemplar by decomposing the exemplar into sentence-level segments and, at each segment, explicitly separating what to say from how to say it. The plan stage extracts question outlines from the source segment, uses short-term memory to resolve pronoun references, and transfers the outlines to the target topic. The adapt stage answers each question with retrieved knowledge, asks the LLM for a verbalized confidence score, refuses answers below threshold $\theta=0.7$, and writes the target segment with long-term memory of prior output to avoid repetition. Across the collected Wikipedia, RoleEE, and USNews pairs, the paper reports that RePA has the lowest hallucinated-token rate, the highest NLI entailment, and the highest adaptive-imitativeness among all compared systems, including retrieval-augmented baselines; ablations attribute the largest factuality drop to removing the outline step and a clear drop to removing the refusal mechanism.

Load-bearing premise

The framework assumes that a verbalized confidence score at or above 0.7 reliably marks an answer as factually correct for the target topic, and a lower score marks the question as better dropped; this confidence-to-correctness link is borrowed from short-answer QA calibration and is not separately calibrated for long expository generation in this paper.

Editorial extensions

If this is right

  • One-shot expository generation becomes practical: a new topic text can be written from a single similar exemplar without domain corpora or fine-tuning.
  • Refusing low-confidence answers to transferred questions cuts hallucinated tokens and raises NLI entailment across all three datasets, so factuality improves by skipping rather than guessing.
  • The question-outline representation is the main carrier of imitation: removing the outline step causes the largest drop in performance of any ablation.
  • The recurrent segment-by-segment setup with short-term input memory and long-term output memory extends the approach to arbitrarily long texts, as shown in the manually curated longer-text case.
  • LLM judges for imitativeness and adaptiveness agree with human judges at roughly 79–83 percent, comparable to human-human agreement, so the new metrics are usable without large annotation budgets.

Reading between the lines

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

  • Beyond the paper, the calibrated-refusal idea is the most portable piece: any plan-then-generate pipeline that transfers questions across topics could adopt the same verbalized-confidence gate, but the $\theta=0.7$ threshold likely needs re-calibration per domain and per backbone LLM.
  • Because retrieval in RePA draws on open-web sources, its factuality advantage may shrink when target-topic facts are absent from the retriever's index; testing RePA on a dataset whose ground truth is not retrievable would separate adaptation skill from retrieval luck.
  • The paper's own error-propagation discussion implies end-to-end gains are limited by the weakest module rather than the average module quality: a Clarify mistake propagates into Outline and then into Calibrated-QA, so component-level ablations understate coupling effects.
  • The same exemplar-based recipe should transfer to structured expository verticals beyond encyclopedia articles, such as product descriptions, faculty profiles, and event pages, provided source and target topics can be paired by embedding similarity; the three datasets already span open-domain and domain-specific settings but not these verticals.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces Exemplar-Based Expository Text Generation, a one-shot task in which a model writes an expository text on a target topic by imitating a single exemplar on a related source topic. The authors propose RePA (Recurrent Plan-then-Adapt), a pipeline that segments the exemplar, clarifies pronouns with a short-term memory, generates transferable question outlines, answers them with retrieval-augmented calibrated QA and a confidence-threshold refusal mechanism, and writes each segment with long-term memory for coherence. They also define two LLM-as-a-judge metrics (Imitativeness and Adaptiveness) plus an F1-style combined score, and collect three datasets (Wikipedia, RoleEE, USNews). Experiments compare RePA against direct prompting, rolling LLM, o1, Self-Refine, and retrieval-augmented variants, with ablations and human-agreement checks for the LLM judge. The central claim is that RePA outperforms strong baselines in producing factual, consistent, and relevant texts for this task.

Significance. If the claims are borne out, the paper makes a useful contribution: it identifies a practical and under-explored task, offers a modular prompting framework that is agnostic to the backbone LLM, releases code and datasets, and provides task-specific evaluation metrics that are checked against human judgments. The strengths include the breadth of the experimental setup (three datasets, two backbone families, ablations, human agreement for the LLM judge), the transparent treatment of some failure modes in Appendix F, and the explicit discussion of limitations. However, the central factual claim rests on an unvalidated confidence-calibration mechanism, and the reported comparisons lack inferential statistics. The paper is therefore a solid step toward a practical recipe, but the evidence as presented is not yet sufficient to support the strong conclusion that RePA 'significantly outperforms' baselines on factuality.

major comments (3)
  1. [§3.2.2, §4.3, Appendix F] The refusal mechanism in Calibrated-QA is load-bearing for the paper's factual claim, but no calibration analysis is reported. The paper borrows verbalized confidence calibration from short-answer QA (Xiong et al., 2023; Tian et al., 2023) and sets θ=0.7, yet does not report the distribution of confidence scores, the fraction of refused questions, or the accuracy of accepted versus refused answers on any of the three datasets. The -F ablation in Table 3 shows only that removing refusal changes metrics (Halluc 6.5714→7.0788, NLI-E 0.7927→0.7481, NLI-C 0.0439→0.0859); it does not demonstrate that the specific threshold 0.7 separates correct from incorrect answers. If verbalized confidence is miscalibrated for these automatically generated, sometimes partially answerable questions, the mechanism will either retain hallucinations or drop valid facts, and Appendix F already admits that retrieval failures cascade into omitted facts. The authors should provide calibration curves or at least a threshold sensitivity analysis, including the accuracy of kept versus refused answers.
  2. [§4.4, Tables 1–3, Appendix E] The central comparison 'RePA significantly outperforms baselines' is not supported by inferential statistics. All reported numbers are point estimates from a single run, with no variance, confidence intervals, or significance tests. Given the scale of differences on some metrics (e.g., ROUGE or Imitativeness), it is unclear which differences are meaningful. In addition, the Halluc metric is underspecified: the paper does not define how hallucinated tokens are identified, whether comparison to 'inputs and ground truths' is token-level or span-level, or how partially correct named-entity substitutions are counted. The NLI-based factuality metrics are human-validated in Appendix E on outputs from only LLM+Retr and RoM, not on RePA outputs, leaving the validity of the metric for the proposed system untested. The authors should report multiple seeds or bootstrap confidence intervals, define Halluc precisely, and extend the human validation to RePA outputs.
  3. [Appendix F, §5.2] The paper acknowledges in Appendix F that Clarify errors, retriever limitations, and QA failures can cascade, leading to omissions of crucial facts and incomplete output, but the only evidence against this is the ablation in Table 3 showing that removing components degrades average metrics. That ablation does not measure how often the refusal mechanism drops valid facts or how often retrieval failure produces incomplete outputs. Since the task requires adapting a large portion of the exemplar's content, a completeness analysis (e.g., proportion of outline questions refused or left unanswered, and the correlation of refusals with human judgments of completeness) is needed to support the adaptiveness and factuality claims.
minor comments (5)
  1. [Table 10 (Appendix F)] The header 'Halluc↓ ↑ NLI-E ↓NLI-C' contains misplaced arrows; it should read 'Halluc↓, NLI-E↑, NLI-C↓' to match the direction of the reported metrics.
  2. [§3.2] In the formal definition of the recurrent unit, 'shot-term memory' appears to be a typo; it should be 'short-term memory'.
  3. [Appendix D.2] The claim that self-enhancement bias does not influence the evaluation is too strong: using the same backbone for all systems removes the most direct form of the bias, but the judge may still systematically prefer stylistic properties of one pipeline over another; the human-agreement study covers only a subset, so the claim should be softened.
  4. [§4.1] The Wikipedia dataset description says 'percentage of common category tags higher than 0.3' but it is ambiguous whether pairs are kept when they share more than 30% of categories or when the divergence is less than 30%; please clarify the filtering direction.
  5. [§4.4] The Halluc metric is introduced without a precise algorithmic definition (e.g., how hallucinated tokens are matched against inputs and ground truths, and how tokenization of named entities is handled); a formal definition or pseudo-code would improve reproducibility.

Circularity Check

1 steps flagged · score 3.0 of 10

Novelty claim rests on an unverified self-citation; the empirical results remain independently supported.

  1. self citation load bearing [Abstract and Contributions list; cf. Section 2.2]
    "Current methods fall short due to their reliance on extensive exemplar data... (1) We are the first to study Exemplar-Based Expository Text Generation task, addressing a practical yet under-explored area with broad applications."

    The 'first to study' premise is justified only by contrasting with Balepur et al. 2023, a prior expository text generation paper coauthored by the present senior author (K.C.-C. Chang). Section 2.2 itself classifies Balepur et al. under 'expository text generation,' and Section 3.1's task definition—given an expository text X on source topic tx, produce Y on ty that imitates X's content/structure while adapting content—is the same single-exemplar imitation-and-adaptation setting the authors now rename 'Exemplar-Based.' The sole stated difference, that prior methods rely on 'extensive domain corpora/exemplar data,' is an assertion about the authors' own prior work, not an externally established fact.

full rationale

The paper's empirical contribution is largely self-contained: RePA is a prompting pipeline with no fitted parameters, and the central factuality/imitativeness/adaptiveness results are measured against external references (ground-truth texts, NLI models, human-validated LLM judges). No equation in the paper defines the output in terms of the evaluation target, and no fitted parameter is relabeled as a prediction. The ablation study (Section 5.2) independently tests each component. The main circularity flag is the novelty framing: the claim to be 'the first to study Exemplar-Based Expository Text Generation' relies on a self-citation (Balepur et al. 2023, shared senior author) whose characterization as relying on 'extensive domain corpora' is not demonstrated; Section 2.2 already lists that same work under 'expository text generation,' making the 'new task' a borderline rename. This affects the contribution claim but not the internal validity of the empirical comparison. Non-circular weaknesses include the unvalidated confidence threshold theta=0.7 borrowed from factoid QA, and the same LLM family serving as both generator and judge—these are correctness risks, not circular reductions. Appendix F's cascading-error rebuttal (every ablation hurts, therefore errors are minimal) is also a weak inference, but it is not a circular step.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the availability of high-quality similar exemplar pairs and on assumptions about LLM behavior and retrieval reliability. Five hand-chosen hyperparameters, notably the confidence threshold of 0.7, directly affect the main factuality results. The paper introduces no new physical or ontological entities; 'Adaptive Imitation' is a task framing concept, not an entity requiring independent evidence.

free parameters (5)
  • confidence threshold theta = 0.7
    Used in Calibrated-QA to filter out low-confidence answers; chosen in Section 4.3 without task-specific calibration evidence, directly controls hallucination versus completeness.
  • topic-pair similarity threshold (Wikipedia) = cosine > 0.95
    Section 4.1 filters topic/text pairs to highly similar pairs, shaping dataset difficulty and generalizability.
  • category common-tag filter threshold (Wikipedia) = > 0.3 (wording ambiguous)
    Section 4.1 uses category overlap to filter pairs; the direction of the filter is unclear as written, but it is a hand-chosen selection threshold.
  • retrieval top-k = top 10 Bing results; top 3 DPR results
    Section 4.3 fixes the amount of external knowledge injected into prompts; affects factuality and output length.
  • generation decoding parameters = temperature 0.3, frequency penalty 0.3, max tokens 256
    Section 4.3 sets these for all models; they are hand-chosen and may affect quality, though they are not the central scientific variable.
assumptions (4)
  • domain assumption A single high-quality exemplar on a similar topic provides a valid structural template for the target topic.
    Stated as a fundamental premise in the Limitations section and enforced through similarity filtering in dataset construction (Section 4.1).
  • domain assumption Verbalized confidence from LLMs is reliable for deciding whether a question is answerable in this retrieval-augmented setting.
    Invoked in Calibrated-QA (Section 3.2.2) with a fixed threshold; only prior short-answer QA calibration results are cited, not task-specific calibration.
  • domain assumption Retrieval sources (Bing web results and the DPR knowledge base) contain the target facts needed to answer transferred outlines.
    Required for the ADAPT stage to produce factual content; the paper discusses retriever limitations in Appendix F.
  • domain assumption The NLI model and LLM-as-a-judge provide valid measurements of factuality, imitativeness, and adaptiveness.
    Section 4.4 and Appendices D-E; human agreement studies cover only subsets, and NLI validation covers only two baselines.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Writing Like the Best: Exemplar-Based Expository Text Generation." pith.science (2026). https://pith.science/paper/VMJ7K5ZP

@misc{pith2026250518859,
  author       = {Pith},
  title        = {Pith review of: Writing Like the Best: Exemplar-Based Expository Text Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VMJ7K5ZP}},
  note         = {Machine review of arXiv:2505.18859}
}
read the original abstract

We introduce the Exemplar-Based Expository Text Generation task, aiming to generate an expository text on a new topic using an exemplar on a similar topic. Current methods fall short due to their reliance on extensive exemplar data, difficulty in adapting topic-specific content, and issues with long-text coherence. To address these challenges, we propose the concept of Adaptive Imitation and present a novel Recurrent Plan-then-Adapt (RePA) framework. RePA leverages large language models (LLMs) for effective adaptive imitation through a fine-grained plan-then-adapt process. RePA also enables recurrent segment-by-segment imitation, supported by two memory structures that enhance input clarity and output coherence. We also develop task-specific evaluation metrics--imitativeness, adaptiveness, and adaptive-imitativeness--using LLMs as evaluators. Experimental results across our collected three diverse datasets demonstrate that RePA surpasses existing baselines in producing factual, consistent, and relevant texts for this task.

Figures

Figures reproduced from arXiv: 2505.18859 by the authors.

Figure 1
Figure 1. An illustration of REPA for Exemplar-Based Expository Text Generation, where yellow text indicates adapted facts and green text indicates discarded facts. can significantly reduce human effort by automat￾ing the writing process, enhance consistency by maintaining a uniform style, and enable scalability by rapidly generating large volumes of text. For expository text generation, prior methods that rely on extensive d… view at source ↗
Figure 2
Figure 2. Overview of REPA. Top left shows the high-level recurrent structure for sequential processing. Bottom details the recurrent unit with a running example: "Clarify" and "Outline" in PLAN, and "Calibrated-QA" (C-QA) and "Write" in ADAPT, with memory usages in "Clarify" (short-term) and "Write" (long-term). input xt as plans for generation. The ADAPT stage is to realize outlines qt flexibly and effectively with knowledg… view at source ↗
Figure 3
Figure 3. A complete running example [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Prompt of the Clarify component in PLAN stage [PITH_FULL_IMAGE:figures/full_fig_p015_4.png]
Figure 5
Figure 5. Figure 5: Prompt of the Outline component in PLAN stage [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Prompt of the Calibrated-QA component in A [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Prompt of the Write component in ADAPT stage [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Prompt of the post-Write summarization step in A [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Prompt of Imitativeness evaluation [PITH_FULL_IMAGE:figures/full_fig_p020_9.png]
Figure 10
Figure 10. Figure 10: Prompt of Adaptiveness evaluation. and reduced adaptiveness. These findings reinforce the limitation of basic generation metrics for this task and highlight the importance of factuality and task-specific evaluation. Though REPA exhibited minor instability in ba￾sic ge…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

68 extracted references · 23 canonical work pages

  1. [1]

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. 2023. Gpt-4 technical report. arXiv preprint arXiv:2303.08774

    arXiv 2023

  2. [2]

    Rilwan Adewoyin, Ritabrata Dutta, and Yulan He. 2022. https://doi.org/10.18653/v1/2022.naacl-main.133 RSTG en: Imbuing fine-grained interpretable control into long- F orm T ext generators . In Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 1822--1835, Se...

    work page doi:10.18653/v1/2022.naacl-main.133 2022

  3. [3]

    Yang Bai, Ziran Li, Ning Ding, Ying Shen, and Hai-Tao Zheng. 2021. Infobox-to-text generation with tree-like planning based attention network. In Proceedings of the Twenty-Ninth International Conference on International Joint Conferences on Artificial Intelligence, pages 3773--3779

    work page 2021

  4. [4]

    Nishant Balepur, Jie Huang, and Kevin Chang. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.729 Expository text generation: Imitate, retrieve, paraphrase . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 11896--11919, Singapore. Association for Computational Linguistics

    doi:10.18653/v1/2023.emnlp-main.729 2023

  5. [5]

    Satanjeev Banerjee and Alon Lavie. 2005. https://aclanthology.org/W05-0909 METEOR : An automatic metric for MT evaluation with improved correlation with human judgments . In Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization , pages 65--72, Ann Arbor, Michigan. Association for Comput...

    2005

  6. [6]

    Iz Beltagy, Matthew E Peters, and Arman Cohan. 2020. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150

    arXiv 2020

  7. [7]

    Bowman, Gabor Angeli, Christopher Potts, and Christopher D

    Samuel R. Bowman, Gabor Angeli, Christopher Potts, and Christopher D. Manning. 2015. https://doi.org/10.18653/v1/D15-1075 A large annotated corpus for learning natural language inference . In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 632--642, Lisbon, Portugal. Association for Computational Linguistics

    doi:10.18653/v1/d15-1075 2015

  8. [8]

    Rebekah Carter, Yenna Salamonson, Lucie M Ramjan, and Elizabeth Halcomb. 2018. Students use of exemplars to support academic writing in higher education: An integrative review. Nurse education today, 65:87--93

    work page 2018

Show all 68 references
  1. [9]

    Jun Chen. 2024. Exploring imitative learning in a blended efl writing class

    2024

  2. [10]

    Albert Gatt and Emiel Krahmer. 2018. Survey of the state of the art in natural language generation: Core tasks, applications and evaluation. Journal of Artificial Intelligence Research, 61:65--170

    2018

  3. [11]

    Seraphina Goldfarb-Tarrant, Tuhin Chakrabarty, Ralph Weischedel, and Nanyun Peng. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.351 Content planning for neural story generation with aristotelian rescoring . In Proceedings of the 2020 Conference on Empirical Methods in Natu...

    2020 doi

  4. [12]

    Jian Guan, Fei Huang, Zhihao Zhao, Xiaoyan Zhu, and Minlie Huang. 2020. https://doi.org/10.1162/tacl_a_00302 A knowledge-enhanced pretraining model for commonsense story generation . Transactions of the Association for Computational Linguistics, 8:93--108

    2020 doi

  5. [13]

    Jian Guan, Xiaoxi Mao, Changjie Fan, Zitao Liu, Wenbiao Ding, and Minlie Huang. 2021. https://doi.org/10.18653/v1/2021.acl-long.499 Long text generation by modeling sentence-level and discourse-level coherence . In Proceedings of the 59th Annual Meeting of the Association for ...

    2021 doi

  6. [14]

    Zhe Hu, Hou Pong Chan, Jiachen Liu, Xinyan Xiao, Hua Wu, and Lifu Huang. 2022. https://doi.org/10.18653/v1/2022.acl-long.163 PLANET : Dynamic content planning in autoregressive transformers for long-form text generation . In Proceedings of the 60th Annual Meeting of the Associ...

    2022 doi

  7. [15]

    Xinyu Hua, Zhe Hu, and Lu Wang. 2019. https://doi.org/10.18653/v1/P19-1255 Argument generation with retrieval, planning, and realization . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2661--2672, Florence, Italy. Association...

    2019 doi

  8. [16]

    Yilun Hua, Zhaoyuan Deng, and Kathleen McKeown. 2023. https://doi.org/10.18653/v1/2023.findings-acl.871 Improving long dialogue summarization with semantic graph representation . In Findings of the Association for Computational Linguistics: ACL 2023, pages 13851--13883, Toront...

    2023 doi

  9. [17]

    Fantine Huot, Joshua Maynez, Shashi Narayan, Reinald Kim Amplayo, Kuzman Ganchev, Annie Priyadarshini Louis, Anders Sandholm, Dipanjan Das, and Mirella Lapata. 2023. Text-blueprint: An interactive platform for plan-based conditional generation. In Proceedings of the 17th Confe...

    2023

  10. [18]

    Maor Ivgi, Uri Shaham, and Jonathan Berant. 2023. https://doi.org/10.1162/tacl_a_00547 Efficient long-text understanding with short-text models . Transactions of the Association for Computational Linguistics, 11:284--299

    2023 doi

  11. [19]

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Ye Jin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12):1--38

    2023

  12. [20]

    Yucheng Jiang, Yijia Shao, Dekun Ma, Sina Semnani, and Monica Lam. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.554 Into the unknown unknowns: Engaged human learning through participation in language model agent conversations . In Proceedings of the 2024 Conference on Emp...

    2024 doi

  13. [21]

    Zhengbao Jiang, Jun Araki, Haibo Ding, and Graham Neubig. 2021. https://doi.org/10.1162/tacl_a_00407 How can we know when language models know? on the calibration of language models for question answering . Transactions of the Association for Computational Linguistics, 9:962--977

    2021 doi

  14. [22]

    Yizhu Jiao, Sha Li, Yiqing Xie, Ming Zhong, Heng Ji, and Jiawei Han. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.395 Open-vocabulary argument role prediction for event extraction . In Findings of the Association for Computational Linguistics: EMNLP 2022, pages 5404--...

    2022 doi

  15. [23]

    Hongye Jin, Xiaotian Han, Jingfeng Yang, Zhimeng Jiang, Zirui Liu, Chia-Yuan Chang, Huiyuan Chen, and Xia Hu. 2024. Llm maybe longlm: Self-extend llm context window without tuning. arXiv preprint arXiv:2401.01325

    2024 arXiv

  16. [24]

    Saurav Kadavath, Tom Conerly, Amanda Askell, Tom Henighan, Dawn Drain, Ethan Perez, Nicholas Schiefer, Zac Hatfield-Dodds, Nova DasSarma, Eli Tran-Johnson, et al. 2022. Language models (mostly) know what they know. arXiv preprint arXiv:2207.05221

    2022 arXiv

  17. [25]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. https://doi.org/10.18653/v1/2020.emnlp-main.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on Empiric...

    2020 doi

  18. [26]

    o ksal, Timo Schick, Anna Korhonen, and Hinrich Sch \

    Abdullatif K \"o ksal, Timo Schick, Anna Korhonen, and Hinrich Sch \"u tze. 2023. Longform: Optimizing instruction tuning for long text generation with corpus extraction. arXiv preprint arXiv:2304.08460

    2023 arXiv

  19. [27]

    Retrieval augmented generation for dialog modeling

    Lilly Kumari, Usama Bin Shafqat, and Nikhil Sarda. Retrieval augmented generation for dialog modeling

  20. [28]

    Gibbeum Lee, Volker Hartmann, Jongho Park, Dimitris Papailiopoulos, and Kangwook Lee. 2023. https://doi.org/10.18653/v1/2023.findings-acl.277 Prompted LLM s as chatbot modules for long open-domain conversation . In Findings of the Association for Computational Linguistics: ACL...

    2023 doi

  21. [29]

    Xiaobo Liang, Zecheng Tang, Juntao Li, and Min Zhang. 2023. https://doi.org/10.18653/v1/2023.acl-long.13 Open-ended long text generation via masked language modeling . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2023 doi

  22. [30]

    Chin-Yew Lin. 2004. https://aclanthology.org/W04-1013 ROUGE : A package for automatic evaluation of summaries . In Text Summarization Branches Out, pages 74--81, Barcelona, Spain. Association for Computational Linguistics

    2004

  23. [31]

    Stephanie Lin, Jacob Hilton, and Owain Evans. 2022. Teaching models to express their uncertainty in words. arXiv preprint arXiv:2205.14334

    2022 arXiv

  24. [32]

    Yuxiang Liu, Jie Huang, and Kevin Chang. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.178 Ask to the point: Open-domain entity-centric question generation . In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 2703--2716, Singapore. Associat...

    2023 doi

  25. [33]

    Junru Lu, Siyu An, Mingbao Lin, Gabriele Pergola, Yulan He, Di Yin, Xing Sun, and Yunsheng Wu. 2023. Memochat: Tuning llms to use memos for consistent long-range open-domain conversation. arXiv preprint arXiv:2308.08239

    2023 arXiv

  26. [34]

    Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. 2024. Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36

    2024

  27. [35]

    Ziming Mao, Chen Henry Wu, Ansong Ni, Yusen Zhang, Rui Zhang, Tao Yu, Budhaditya Deb, Chenguang Zhu, Ahmed Awadallah, and Dragomir Radev. 2022. https://doi.org/10.18653/v1/2022.acl-long.118 DYLE : Dynamic latent extraction for abstractive long-input summarization . In Proceedi...

    2022 doi

  28. [36]

    Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.741 FA ct S core: Fine-grained atomic evaluation of factual precision in long form text generatio...

    2023 doi

  29. [37]

    Piotr Mirowski, Kory W Mathewson, Jaylen Pittman, and Richard Evans. 2023. Co-writing screenplays and theatre scripts with language models: Evaluation by industry professionals. In Proceedings of the 2023 CHI Conference on Human Factors in Computing Systems, pages 1--34

    2023

  30. [38]

    Amit Moryossef, Yoav Goldberg, and Ido Dagan. 2019. https://doi.org/10.18653/v1/N19-1236 S tep-by-step: S eparating planning from realization in neural data-to-text generation . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computa...

    2019 doi

  31. [39]

    Shashi Narayan, Joshua Maynez, Reinald Kim Amplayo, Kuzman Ganchev, Annie Louis, Fantine Huot, Anders Sandholm, Dipanjan Das, and Mirella Lapata. 2023. https://doi.org/10.1162/tacl_a_00583 Conditional generation with a question-answering blueprint . Transactions of the Associa...

    2023 doi

  32. [40]

    Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. 2002. https://doi.org/10.3115/1073083.1073135 B leu: a method for automatic evaluation of machine translation . In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311--3...

    2002

  33. [41]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023 a . https://doi.org/10.1162/tacl_a_00605 In-context retrieval-augmented language models . Transactions of the Association for Computational Linguistics, 11:1316--1331

    2023 doi

  34. [42]

    Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and Yoav Shoham. 2023 b . In-context retrieval-augmented language models. Transactions of the Association for Computational Linguistics, 11:1316--1331

    2023

  35. [43]

    Vipula Rawte, Amit Sheth, and Amitava Das. 2023. A survey of hallucination in large foundation models. arXiv preprint arXiv:2309.05922

    2023 arXiv

  36. [44]

    Yijia Shao, Yucheng Jiang, Theodore Kanell, Peter Xu, Omar Khattab, and Monica Lam. 2024 a . https://doi.org/10.18653/v1/2024.naacl-long.347 Assisting in writing W ikipedia-like articles from scratch with large language models . In Proceedings of the 2024 Conference of the Nor...

    2024 doi

  37. [45]

    Yijia Shao, Yucheng Jiang, Theodore A Kanell, Peter Xu, Omar Khattab, and Monica S Lam. 2024 b . Assisting in writing wikipedia-like articles from scratch with large language models. arXiv preprint arXiv:2402.14207

    2024 arXiv

  38. [46]

    Dinghan Shen, Asli Celikyilmaz, Yizhe Zhang, Liqun Chen, Xin Wang, Jianfeng Gao, and Lawrence Carin. 2019. https://doi.org/10.18653/v1/P19-1200 Towards generating long and coherent text with multi-level latent variable models . In Proceedings of the 57th Annual Meeting of the ...

    2019 doi

  39. [47]

    Weijia Shi, Sewon Min, Michihiro Yasunaga, Minjoon Seo, Rich James, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. 2023. Replug: Retrieval-augmented black-box language models. arXiv preprint arXiv:2301.12652

    2023 arXiv

  40. [48]

    Aviv Slobodkin, Omer Goldman, Avi Caciularu, Ido Dagan, and Shauli Ravfogel. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.220 The curious case of hallucinatory (un)answerability: Finding truths in the hidden states of over-confident large language models . In Proceedings ...

    2023 doi

  41. [49]

    Xiaofei Sun, Zijun Sun, Yuxian Meng, Jiwei Li, and Chun Fan. 2022. https://aclanthology.org/2022.coling-1.556 Summarize, outline, and elaborate: Long-text generation via hierarchical supervision from extractive summaries . In Proceedings of the 29th International Conference on...

    2022

  42. [50]

    Katherine Tian, Eric Mitchell, Allan Zhou, Archit Sharma, Rafael Rafailov, Huaxiu Yao, Chelsea Finn, and Christopher Manning. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.330 Just ask for calibration: Strategies for eliciting calibrated confidence scores from language mod...

    2023 doi

  43. [51]

    Chintalapalli Vijayakumar. 2024. Exemplification in student essay writing: A study of learner corpus of essay writing (lcew). International Journal of Applied Linguistics, 34(4):1514--1532

    2024

  44. [52]

    Qingyue Wang, Liang Ding, Yanan Cao, Zhiliang Tian, Shi Wang, Dacheng Tao, and Li Guo. 2023. Recursively summarizing enables long-term dialogue memory in large language models. arXiv preprint arXiv:2308.15022

    2023 arXiv

  45. [53]

    Rosemary Wette. 2014. Teachers' practices in eap writing instruction: Use of models and modeling. System, 42:60--69

    2014

  46. [54]

    Zhiwei Wu. 2019. Understanding students’ mimicry, emulation and imitation of genre exemplars: An exploratory study. English for Specific Purposes, 54:127--138

    2019

  47. [55]

    Miao Xiong, Zhiyuan Hu, Xinyang Lu, Yifei Li, Jie Fu, Junxian He, and Bryan Hooi. 2023. Can llms express their uncertainty? an empirical evaluation of confidence elicitation in llms. arXiv preprint arXiv:2306.13063

    2023 arXiv

  48. [56]

    Han Xu, Xingyuan Wang, and Haipeng Chen. 2024. Towards real-time and personalized code generation. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 5568--5569

    2024

  49. [57]

    Ikuya Yamada, Akari Asai, Jin Sakuma, Hiroyuki Shindo, Hideaki Takeda, Yoshiyasu Takefuji, and Yuji Matsumoto. 2020. https://doi.org/10.18653/v1/2020.emnlp-demos.4 W ikipedia2 V ec: An efficient toolkit for learning and visualizing the embeddings of words and entities from W i...

    2020 doi

  50. [58]

    Kevin Yang, Dan Klein, Nanyun Peng, and Yuandong Tian. 2023 a . https://doi.org/10.18653/v1/2023.acl-long.190 DOC : Improving long story coherence with detailed outline control . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume...

    2023 doi

  51. [59]

    Kevin Yang, Yuandong Tian, Nanyun Peng, and Dan Klein. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.296 Re3: Generating longer stories with recursive reprompting and revision . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, page...

    2022 doi

  52. [60]

    Yuqing Yang, Ethan Chern, Xipeng Qiu, Graham Neubig, and Pengfei Liu. 2023 b . Alignment for honesty. arXiv preprint arXiv:2312.07000

    2023 arXiv

  53. [61]

    Wang You, Wenshan Wu, Yaobo Liang, Shaoguang Mao, Chenfei Wu, Maosong Cao, Yuzhe Cai, Yiduo Guo, Yan Xia, Furu Wei, et al. 2023. Eipe-text: Evaluation-guided iterative plan extraction for long-form narrative text generation. arXiv preprint arXiv:2310.08185

    2023 arXiv

  54. [62]

    Hanning Zhang, Shizhe Diao, Yong Lin, Yi R Fung, Qing Lian, Xingyao Wang, Yangyi Chen, Heng Ji, and Tong Zhang. 2023. R-tuning: Teaching large language models to refuse unknown questions. arXiv preprint arXiv:2311.09677

    2023 arXiv

  55. [63]

    Ruqing Zhang, Jiafeng Guo, Yixing Fan, Yanyan Lan, and Xueqi Cheng. 2019. Outline generation: Understanding the inherent content structure of documents. In Proceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 745--754

    2019

  56. [64]

    Yusen Zhang, Ansong Ni, Ziming Mao, Chen Henry Wu, Chenguang Zhu, Budhaditya Deb, Ahmed Awadallah, Dragomir Radev, and Rui Zhang. 2022. https://doi.org/10.18653/v1/2022.acl-long.112 S umm ^n : A multi-stage summarization framework for long input dialogues and documents . In Pr...

    2022 doi

  57. [65]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2024. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36

    2024

  58. [66]

    Wangchunshu Zhou, Yuchen Eleanor Jiang, Peng Cui, Tiannan Wang, Zhenxin Xiao, Yifan Hou, Ryan Cotterell, and Mrinmaya Sachan. 2023. Recurrentgpt: Interactive generation of (arbitrarily) long text. arXiv preprint arXiv:2305.13304

    2023 arXiv

  59. [67]

    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...

  60. [68]

    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 gl...

Pith tools

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