Pith. sign in

REVIEW 4 major objections 5 minor 46 references

STORYTELLER: An Enhanced Plot-Planning Framework for Coherent and Cohesive Story Generation

T0 review · 4 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 STORYTELLER claims that dynamic plot nodes built from subject-verb-object triples make long-form AI stories cohere, beating baselines in 84.33% of human preference comparisons.

desk verdict The architecture is a genuine integration advance, but the headline win rate is confounded by output length and by using GPT-4o as both generator and judge; the evaluation needs to be redone before the claims hold. read the letter →

arxiv 2506.02347 v1 pith:S4QD3KV3 submitted 2025-06-03 cs.CL

Jiaming Li , Yukun Chen , Ziqiang Liu , Minghuan Tan , Lei Zhang , Yunshui Li , Run Luo , Longze Chen
show 5 more authors
Jing Luo Ahmadreza Argha Hamid Alinejad-Rokny Wei Zhou Min Yang
This is my paper · ORCID
classification cs.CL
keywords storygenerationplotplanningSVOtripletsknowledgegraphnarrativecoherencelong-formtexthumanpreferenceevaluationLLM
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

The paper argues that long-form story generation fails because plot events are planned once and then written independently, so later chapters drift from earlier ones. STORYTELLER tries to fix this by representing every important event as a subject-verb-object node, storing the nodes in a dated storyline and a character/entity knowledge graph, and consulting both before each new plot node is accepted. On 200 Reddit writing prompts, human preference evaluation gives STORYTELLER an 84.33% average win rate over GPT-4o, Qwen2-72B, Llama-3.1-70B, LongWriter, and DOC v2, with the largest edge on relevance and coherence. If the result holds, it means a small structured planning layer can keep LLMs on-plot across thousands of words, which is the main obstacle to automated novel-length fiction.

What carries the argument

The load-bearing object is the plot node: a single event written as <subject, verb, object>, or <subject, verb, subject> for intransitive actions. Nodes are organized into a storyline that timestamps every event and a narrative entity knowledge graph that stores characters, locations, and their relations. During generation, a proposed pseudo-CPN is compared against the most recent events from the storyline and the nodes related to it in the graph; an LLM accepts it or rewrites it before it becomes the next plot node. This review loop is what forces each plot step to be grounded in what has already happened, which is the mechanism the paper credits for coherence and relevance.

What would settle it

Generate the same 200 prompts with STORYTELLER and its ablations constrained to match GPT-4o's average length of about 1,076 words; if the win rate drops toward chance, length carries the result rather than the plot-node mechanism. A cleaner version reruns the human preference study with a style-matched, length-matched baseline.

Watch

Extended reading notes

Core claim

The central claim is that continuous, bidirectional interaction between a plot plan and the generating model is what creates coherence, not a better generator or more planning at the outline level. STORYTELLER turns each story event into an SVO/SVS triplet node, classifies nodes as chapter-begin, chapter-plot, or chapter-end, and uses a review step that checks each proposed plot node against recent storyline events and related knowledge-graph entities, modifying it when it conflicts. The paper reports this pipeline beats every baseline in side-by-side human preference, winning 91% against GPT-4o and 82% against DOC v2, and that ablations—dropping the knowledge graph lowers relevance by 10.3 points; dropping the pseudo-CPN review lowers the overall score by 10.3—confirm both modules contribute.

Load-bearing premise

The observed quality advantage may come from STORYTELLER's much longer, GPT-4o-styled output rather than from the plot-node modules, because the evaluation prompt demands at least 3,000 words and maximal narrative expansion.

Editorial extensions

If this is right

  • STORYTELLER's pipeline can be applied on top of any instruction-tuned LLM, since the modules only prompt the generator rather than retraining it.
  • Longer, more detailed plot-node plans should let authors guide stories at a chapter level—specifying a beginning and ending node effectively sets a contract the generation must satisfy.
  • The narrative entity knowledge graph makes the system extensible: new characters or subplots can be introduced by adding graph nodes, and the review step will surface connections to existing entities.
  • The ablation numbers imply that the plot review step, not just the outline, is responsible for most of the gain; removing it drops the overall score by 10.3 and average length by 2,390 words.

Reading between the lines

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

  • The 7,594-word average output and the explicit 'at least 3,000 words, maximal narrative expansion' evaluation instruction mean part of the win could be a length premium; a fair test would compare at matched lengths.
  • Because GPT-4o is both the generation engine and the judge, the evaluation may inherit GPT-4o's stylistic preferences; human judges' strong agreement with that preference suggests the effect is real but the relative contributions of style and structure remain open.
  • The SVO review loop resembles a constrained-decoding or state-tracking approach; one testable extension is to run the same loop on existing hierarchical generators and see whether the gain transfers.
  • The knowledge graph could be replaced by a cheaper vector store of event embeddings, which would test whether the graph's relational structure is the active ingredient.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper introduces STORYTELLER, a three-stage plot-planning framework for long-form story generation. The method builds a plot node representation from SVO/SVS triplets, maintains a storyline with time-stamped nodes, and uses a narrative entity knowledge graph (NEKG) to retrieve and review pseudo plot nodes during generation. The authors evaluate STORYTELLER on 200 WritingPrompts prompts against GPT-4o, Qwen2-72B, Llama-3.1-70B, LongWriter variants, and DOC v2, reporting an Overall score of 89.4 and a human-preference win rate of 84.33% average. Ablations remove the NEKG and the Pseudo CPN Review module. The central claim is that the proposed modules, rather than the underlying LLM or longer output, drive the observed quality advantage.

Significance. If the central claim were cleanly established, STORYTELLER would be a notable contribution to long-form story generation: the plot-node/SVO mechanism, the dynamic storyline, and the knowledge-graph interaction are well-motivated and described in enough detail to be reproduced. The inclusion of human preference evaluation, inter-annotator agreement (Cohen's kappa), and a documented prompt suite are strengths. However, the paper's headline claims ('significantly outperforms', 'state-of-the-art') rest on an evaluation whose validity is weakened by a major length confound, by the fact that GPT-4o is simultaneously the generation engine for the proposed method and the judge for all automated evaluations, and by an empty human-evaluation template that prevents verification of what human raters were shown. The diversity metric DistinctL-n is also introduced without validation. These issues are fixable in principle, but they currently prevent the empirical contribution from being assessed.

major comments (4)
  1. [§4.1, §4.2, Table 3, Appendix C.1–C.2] The reported win rates and overall scores are confounded with output length. Table 3 shows STORYTELLER averaging 7,594 words versus 1,076 for GPT-4o and 5,204 for the no-Pseudo-CPN-Review ablation; Appendix C.1 instructs all generators to produce stories 'at least 3,000 words long' and to 'prioritize maximal narrative expansion', and Appendix C.2 gives the same length instruction to the GPT-4o judge. A judge told that longer is better will likely rank the 7.6k-word output above the 1.1k-word GPT-4o output regardless of plot-node structure. The paper does not report length-matched comparisons or length as a covariate, so the central claim that the observed gains come from plot nodes, STORYLINE, and NEKG is not supported.
  2. [§4.1 Configuration, §4.2, §4.3, Appendix C.3] The evaluation uses GPT-4o as both the generation engine for STORYTELLER (and DOC v2) and as the judge for all automated metrics and preferences. This creates a possible same-model bias that is acknowledged nowhere in the paper. More importantly, the human-preference evaluation, which is the only independent check, is not verifiable: Appendix C.3, titled 'Human Evaluation System', contains only a heading 'HumanEval Template' and no actual template, instructions, or example. Without knowing what human raters were asked (in particular, whether they were given length-neutral instructions), the reported 91% human win rate against GPT-4o cannot be interpreted as evidence for the proposed modules rather than for length or style.
  3. [Table 3, Figures 4–5, §4.2–§4.3] No error bars, confidence intervals, or significance tests are reported for any of the headline numbers. The Overall scores in Table 3 (e.g., 89.4 for STORYTELLER versus 62.7 for LongWriter-glm4-9b) and the preference percentages in Figures 4 and 5 are presented as exact point estimates, but the paper uses the word 'significantly' throughout. The human-preference evaluation also does not state the number of pairwise comparisons, the number of annotators per pair beyond 'at least two', or the sampling of stories. These omissions make it impossible to assess whether the differences are reliable.
  4. [§4.1, §4.5, Eq. (3)] The DistinctL-n metric introduced in Eq. (3) multiplies the distinct-n-gram ratio by (1 + log(word_count)). This is an ad hoc length normalization that systematically inflates the score of longer texts; since STORYTELLER generates much longer outputs than most baselines, the claimed diversity advantage may be an artifact of the metric rather than of genuine lexical diversity. The paper does not validate DistinctL-n against any established diversity measure or report results with the unnormalized Distinct-n. The diversity claim in Section 4.5 therefore needs reanalysis or the metric needs justification.
minor comments (5)
  1. [Table 2] The caption says 'Statiscs of WRITINGPROMPTSdataset'; fix the typo to 'Statistics'.
  2. [§4.3] The sentence 'GPT-4o is utilized as the annotato' should read 'annotator'.
  3. [§4.5] The phrase 'Textual diversity consititues' contains a typo; it should be 'constitutes'.
  4. [Appendix C.2] The checklist text says 'this facter' and later 'as your final verdict with a label'; the former should be 'factor' and the passage should be polished for clarity.
  5. [§3.3] The sentence 'To address the, Pseudo CPNis generated...' seems to be missing a noun after 'the'; it should say 'To address this challenge' or similar.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity: STORYTELLER's headline SOTA claim is grounded in independent human preference evaluation; the only self-definitional element is the peripheral DistinctL-n diversity metric, which rewards length by construction.

  1. self definitional [Section 4.1, Equation 3; Table 5]
    "DistinctL-n, defined in Equation 3, incorporates text length to provide a fairer assessment of diversity in longer texts. DistinctL−n = unique n-grams total n-grams ×(1 + log(word_count))(3)"

    The metric used to support the diversity claim is defined with a built-in length multiplier. STORYTELLER outputs 7,594 words versus GPT-4o's 1,076 words (Table 3), so its higher DistinctL-4/5 scores in Table 5 are partly forced by the formula itself, independent of underlying n-gram variety. This is a self-definitional artifact of the reported diversity evidence, though it is not central to the main SOTA claim, which is grounded by independent human preference evaluation.

full rationale

STORYTELLER's central claims are empirical: an 84.33% average human-preference win rate and an 89.4 Overall score from human and LLM-judge comparisons against external baselines. Those claims are not derived from the model's own equations or from a chain of self-citations; the evaluation uses independent human raters and standard external baselines (GPT-4o, Qwen2, Llama, LongWriter, DOC v2). Nothing in the framework defines 'winning' in terms of the model's own parameters, and the headline result does not reduce to a fitted value or to an imported uniqueness theorem. The authors' self-citations (e.g., Zhang et al. 2023) are incidental, not load-bearing. One edge: the DistinctL-n metric in Eq. 3 explicitly multiplies a distinct-n ratio by (1 + log(word_count)), so the diversity comparison partly rewards length by construction; but the paper's central SOTA claim does not rest on this metric. The length/style confounds in the evaluation protocol are important validity risks, but they are concerns about experimental control, not circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 2 invented entities

The method is prompt-based and has no trained parameters; the main inputs are domain assumptions about LLM reliability, evaluation validity, length neutrality, and dataset representativeness. The hand-designed DistinctL-n metric is a free parameter, and the two core modules (STORYLINE, NEKG) are new components without external evidence beyond the paper's own experiments.

free parameters (1)
  • DistinctL-n length normalization factor
    Eq. 3 uses (1 + log(word_count)) as a length correction; it is hand-chosen and not derived from data, and it favors long outputs when reporting diversity.
assumptions (4)
  • domain assumption LLM-generated SVO plot nodes and LLM-based CPN review are reliable enough to maintain coherence
    The whole framework (Sections 3.2-3.4) assumes the engine can extract and review NODES without verification; the paper does not report extraction accuracy or review failure rates.
  • domain assumption GPT-4o is an unbiased judge of stories, including stories it generated itself
    Section 4.1 states GPT-4o serves as judge for all evaluations while also being the engine for STORYTELLER and DOC v2; no control for self-preference is described.
  • domain assumption Story length does not dominate preference and quality judgments
    STORYTELLER averages 7,594 words vs 1,076 for GPT-4o and 5,204 for its w/o Pseudo CPN Review ablation (Table 3); the evaluation prompt (Appendix C.1) explicitly asks for at least 3,000 words and maximal expansion, so length is a confound.
  • domain assumption The 200 selected [WP] prompts are representative of WritingPrompts story generation
    Section 4.1 says 200 prompts were randomly selected from WritingPrompts, but no distribution or representativeness analysis is given; results may not generalize to other tags or genres.
invented entities (2)
  • STORYLINE
    purpose: Ordered store of plot nodes with timestamps, used to retrieve recent events during generation
    Introduced by this paper as a core module; its benefit is only supported by the paper's own ablation, with no external validation.
  • NEKG (narrative entity knowledge graph)
    purpose: Graph of entities and relations queried with Cypher to supply context for CPN review
    The paper implements it in Neo4j and claims it improves relevance and coherence, but provides no independent evidence of graph correctness or retrieval quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of STORYTELLER: An Enhanced Plot-Planning Framework for Coherent and Cohesive Story Generation." pith.science (2026). https://pith.science/paper/S4QD3KV3

@misc{pith2026250602347,
  author       = {Pith},
  title        = {Pith review of: STORYTELLER: An Enhanced Plot-Planning Framework for Coherent and Cohesive Story Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4QD3KV3}},
  note         = {Machine review of arXiv:2506.02347}
}
read the original abstract

Stories are central to human culture, serving to share ideas, preserve traditions, and foster connections. Automatic story generation, a key advancement in artificial intelligence (AI), offers new possibilities for creating personalized content, exploring creative ideas, and enhancing interactive experiences. However, existing methods struggle to maintain narrative coherence and logical consistency. This disconnect compromises the overall storytelling experience, underscoring the need for substantial improvements. Inspired by human cognitive processes, we introduce Storyteller, a novel approach that systemically improves the coherence and consistency of automatically generated stories. Storyteller introduces a plot node structure based on linguistically grounded subject verb object (SVO) triplets, which capture essential story events and ensure a consistent logical flow. Unlike previous methods, Storyteller integrates two dynamic modules, the STORYLINE and narrative entity knowledge graph (NEKG),that continuously interact with the story generation process. This integration produces structurally sound, cohesive and immersive narratives. Extensive experiments demonstrate that Storyteller significantly outperforms existing approaches, achieving an 84.33% average win rate through human preference evaluation. At the same time, it is also far ahead in other aspects including creativity, coherence, engagement, and relevance.

Figures

Figures reproduced from arXiv: 2506.02347 by the authors.

Figure 1
Figure 1. Story generation still faces challenges like [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of STORYTELLER. The method is divided into three stages. This figure illustrates the complete process by which STORYTELLER transforms user input into the final story. Additionally, it highlights the information interaction between various stages and both the STORYLINE and NEKG module. ing, which enables task execution based solely on prompt content without requiring parameter up￾dates (Dong et al., 2022; We… view at source ↗
Figure 3
Figure 3. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Results of preference evaluation using GPT [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Results of human preference evaluation. STO￾RYTELLER achieves outstanding preference wins, with 79% against LongWriter-llama3.1-8b, over 80% against other models, including 91% against GPT-4o and 83% against LongWriter-glm4-9b. word count than Llama3.1-70B-Instruct; ho…
Figure 6
Figure 6. Figure 6: Proportion of CPN_Review Types. The fig [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 22 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]

    AI@Meta. 2024. https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md Llama 3 model card

    2024

  3. [3]

    Amal Alabdulkarim, Siyan Li, and Xiangyu Peng. 2021. Automatic story generation: Challenges and attempts. arXiv preprint arXiv:2102.12634

    work page Pith review arXiv 2021

  4. [4]

    Arwa I Alhussain and Aqil M Azmi. 2021. Automatic story generation: A survey of approaches. ACM Computing Surveys (CSUR), 54(5):1--38

    work page 2021

  5. [5]

    Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng X...

    arXiv 2023

  6. [6]

    Yushi Bai, Jiajie Zhang, Xin Lv, Linzhi Zheng, Siqi Zhu, Lei Hou, Yuxiao Dong, Jie Tang, and Juanzi Li. 2024. Longwriter: Unleashing 10,000+ word generation from long context llms. arXiv preprint arXiv:2408.07055

    arXiv 2024

  7. [7]

    Dorothea Brande. 2013. from becoming a writer. In Creative Writing, pages 424--427. Routledge

    work page 2013

  8. [8]

    Tom B Brown. 2020. Language models are few-shot learners. arXiv preprint arXiv:2005.14165

    arXiv 2020

Show all 46 references
  1. [9]

    Roy Peter Clark. 2008. Writing tools: 55 essential strategies for every writer. Little, Brown Spark

    2008

  2. [10]

    William Cook. 2011. PLOTTO: the master book of all plots. Tin House Books

    2011

  3. [11]

    Don Cupitt. 1991. What is a Story? Trinity PressIntl

    1991

  4. [12]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/v1/N19-1423 BERT : Pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associat...

    2019 doi

  5. [13]

    Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Zhiyong Wu, Baobao Chang, Xu Sun, Jingjing Xu, and Zhifang Sui. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234

    2022 arXiv

  6. [14]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  7. [15]

    Angela Fan, Mike Lewis, and Yann Dauphin. 2018. Hierarchical neural story generation. In Conference of the Association for Computational Linguistics (ACL)

    2018

  8. [16]

    Angela Fan, Mike Lewis, and Yann Dauphin. 2019. https://doi.org/10.18653/v1/P19-1254 Strategies for structuring story generation . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2650--2660, Florence, Italy. Association for Com...

    2019 doi

  9. [17]

    Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie H...

    2024 arXiv

  10. [18]

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

    2020

  11. [19]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. 2023. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating S...

    2023

  12. [20]

    Anne Lamott. 1995. Bird by bird: Some instructions on writing and life. Vintage

    1995

  13. [21]

    Yoav Levine, Noam Wies, Daniel Jannai, Dan Navon, Yedid Hoshen, and Amnon Shashua. 2022. https://openreview.net/forum?id=lnEaqbTJIRz The inductive bias of in-context learning: Rethinking pretraining example design . In International Conference on Learning Representations

    2022

  14. [22]

    Boyang Li, Stephen Lee-Urban, George Johnston, and Mark Riedl. 2013. Story generation with crowdsourced plot graphs. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 27, pages 598--604

    2013

  15. [23]

    Jiwei Li, Michel Galley, Chris Brockett, Jianfeng Gao, and Bill Dolan. 2016. https://doi.org/10.18653/v1/N16-1014 A diversity-promoting objective function for neural conversation models . In Proceedings of the 2016 Conference of the North A merican Chapter of the Association f...

    2016 doi

  16. [24]

    Tianle Li, Wei-Lin Chiang, Evan Frick, Lisa Dunlap, Tianhao Wu, Banghua Zhu, Joseph E Gonzalez, and Ion Stoica. 2024. From crowdsourced data to high-quality benchmarks: Arena-hard and benchbuilder pipeline. arXiv preprint arXiv:2406.11939

    2024 arXiv

  17. [25]

    Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. 2023. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9):1--35

    2023

  18. [26]

    James R Meehan. 1977. Tale-spin, an interactive program that writes stories. In Ijcai, volume 77, pages 91--98

    1977

  19. [27]

    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

  20. [28]

    Hello gpt-4o | openai

    OpenAI. Hello gpt-4o | openai. https://openai.com/index/hello-gpt-4o/

  21. [29]

    Xiangyu Peng, Siyan Li, Sarah Wiegreffe, and Mark Riedl. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.520 Inferring the reader: Guiding automated story generation with commonsense reasoning . In Findings of the Association for Computational Linguistics: EMNLP 2022, pa...

    2022 doi

  22. [30]

    Xiangyu Peng, Kaige Xie, Amal Alabdulkarim, Harshith Kayam, Samihan Dani, and Mark O Riedl. 2021. Guiding neural story generation with reader models. arXiv preprint arXiv:2112.08596

    2021 arXiv

  23. [31]

    Hannah Rashkin, Asli Celikyilmaz, Yejin Choi, and Jianfeng Gao. 2020. Plotmachines: Outline-conditioned generation with dynamic plot state tracking. arXiv preprint arXiv:2004.14967

    2020 arXiv

  24. [32]

    Brandon Tearse, Noah Wardrip-Fruin, and Michael Mateas. 2010. Minstrel remixed: Procedurally generating stories. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment, volume 6, pages 192--197

    2010

  25. [33]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, et al. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    2023 arXiv

  26. [34]

    Scott R Turner. 1993. Minstrel: a computer model of creativity and storytelling. University of California, Los Angeles

    1993

  27. [35]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in neural information processing systems, 30

    2017

  28. [36]

    Yuxin Wang, Jieru Lin, Zhiwei Yu, Wei Hu, and B \"o rje F Karlsson. 2023. Open-world story generation with structured knowledge enhancement: A comprehensive survey. Neurocomputing, page 126792

    2023

  29. [37]

    Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, and William Fedus. 2022. https://openreview.net/forum?id=yzkSU5zd...

    2022

  30. [38]

    Kevin Yang, Dan Klein, Nanyun Peng, and Yuandong Tian. 2022 a . Doc: Improving long story coherence with detailed outline control. arXiv preprint arXiv:2212.10077

    2022 arXiv

  31. [39]

    Kevin Yang, Yuandong Tian, Nanyun Peng, and Dan Klein. 2022 b . Re3: Generating longer stories with recursive reprompting and revision. arXiv preprint arXiv:2210.06774

    2022 arXiv

  32. [40]

    Lili Yao, Nanyun Peng, Ralph Weischedel, Kevin Knight, Dongyan Zhao, and Rui Yan. 2019. Plan-and-write: Towards better automatic storytelling. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 7378--7385

    2019

  33. [41]

    Lei Zhang, Yunshui Li, Ziqiang Liu, Jiaxi yang, Junhao Liu, and Min Yang. 2023. https://arxiv.org/abs/2312.09542 Marathon: A race through the realm of long context with large language models . Preprint, arXiv:2312.09542

    2023 arXiv

  34. [42]

    Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. 2022. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

    2022 arXiv

  35. [43]

    Zihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. 2021. Calibrate before use: Improving few-shot performance of language models. In International conference on machine learning, pages 12697--12706. PMLR

    2021

  36. [44]

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. 2024. Lima: Less is more for alignment. Advances in Neural Information Processing Systems, 36

    2024

  37. [45]

    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...

  38. [46]

    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.